ambean/construct-validity-review · Datasets at Hugging Face

liuRepoBenchBenchmarkingRepositorylevel2024

RepoBench: Benchmarking Repository-Level Code Auto-Completion Systems

Include

null

Evaluating whether models can do repository-level code generation.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Repository-level coding tasks.

No

null

Comprehensive

null

Three sub-tasks. (1) the ability to retrieve the most relevant code snippets from a repo (2) predict the next line of code and (3) do both simultaneously.

Retrieval: identify the most relevant code snippets to predict the next line given an in-file context. Generation: predict the next line of code based on a given in-file context.

null

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

1669

Yes

Task, difficulty level, different code masking settings

Random sample (creators defined a task space and sampled from it), Specific criteria (items were taken from a larger set based on specified rules)

Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF)

Not the best. Most programming benchmarks use some kind of execution base metric. Not applicable for the retrieval task here but would be applicable for the generation task.

null

Academia

Yes

null

Test, Train

25301

null

Simple Mean

Yes

Task, programming language,

accuracy@k

https://github.com/Leolty/repobench

REPOBENCH

Contested

Yes

No

No comparisons made

The benchmark is itself realistic

No

null

mean,

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

No

Code Generation

null

['Real task', 'Author-crafted']

['Random', 'Criterion']

['Structured']

['Exact match', 'Soft match']

['Contested']

['Yes']

['Realistic']

['No']

['Partial', 'Representative']

['Mean']

zhangBenchmarkingDataScience2024

Benchmarking Data Science Agents

Include

null

This paper introduces DSEval, a novel bencmark for evaluating LLMs as data science agents.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Data science

Yes

Executing a wide array of data-centric tasks, including manipulation, aggregation, visualization, and analysis, through natural language commands.

Comprehensive

null

The task involves responding to a natural language data science query using both the query itself and a stateful runtime session, which provides contextual information such as variables, execution history, and files. The agent must generate and return executable code that solves the query.

Each problemset is represented as a Python (*.py) file with YAML syntax inside to “configure” the problem, including the query, validator configurations, execution restrictions, and external data required.

null

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), LLM-generated task examples (e.g. Filtered from responses to a prompt)

825

Yes

Difficulty

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

The “Pass Rate”, which is the number of problems passed divided by all problems in the benchmark, is the default metric used to assess the quality of an agent.

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Yes

By subtask dataset

null

https://github.com/MetaCopilot/dseval/tree/master

DSEval

Widely-agreed

Yes

No

Yes

No

No comparisons made

The benchmark is itself realistic

No

Somewhat

While not explicitly addressed, the dataset construction ensures that the task items are sourced from real-world data science tasks -- which adds to its construct validity.

null

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

No

Code Generation

null

['Real task', 'Author-crafted', 'LLM-generated']

['Targeted', 'Criterion']

['Structured']

['Exact match']

['Widely-agreed']

['Yes']

['No']

['Realistic']

['Yes']

['Partial']

null

kimFANToMBenchmarkStresstesting2023

FANToM: A Benchmark for Stress-testing Machine Theory of Mind in Interactions

Include

null

FANTOM is a comprehensive benchmark with 10,000 questions designed to evaluate theory-of-mind reasoning capabilities in LLMs by tracking beliefs across multi-party conversations with information asymmetry. The benchmark reveals that even leading LLMs struggle significantly with belief tracking compared to humans, demonstrating only "illusory" theory-of-mind abilities that break down when tested systematically.

(1) First conversation-based theory-of-mind benchmark, (2) explicit test for consistency checks across diverse question formats, (3) 256 multiparty conversations around a certain topic

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Theory of mind, conversational question-answering

Yes

The goal of FANTOM is to effectively measure how well models can track the belief of multiple characters in conversations where some information may be inaccessible to some participants. Our aim is to design questions at different levels that evaluate a model’s capability for a coherent understanding of others’ mental states. In doing so, we are particularly interested in identifying instances of illusory ToM, which we define as situations where a model may answer some questions correctly but fails to answer others that require the same type of ToM reasoning.

Subset

null

Models read a multiparty conversation (short or full) and answer six types of questions (free‑form, multiple‑choice, list, yes/no) about participants’ beliefs or answerability. Model responses are expected to be "yes", "knows", "does know", or similar type of answers.

null

Crowd-sourced task examples (e.g. Prolific-created tasks), LLM-generated task examples (e.g. Filtered from responses to a prompt)

f 256 conversations with 1,415 (Theory of mind) belief questions, 703 fact-checking questions, and 2689 answerability questions.

No

belief questions (theory of mind; how the characters beliefs of something), fact-checking questions, answerability questions

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics)

BELIEFQ[Dist.] judged by sentence‑BERT cosine vs. two references and token‑F1; others via strict match

Conversations autogenerated with davinci‑003, inaccessible info & QA generated with GPT‑4, all sets manually validated by 32 MTurk workers

Academia

Yes

https://hyunw.kim/fantom/

null

Test

null

Open text, binary or multiple choice types

Simple Mean

Yes

Sub‑scores are broken out by each question family — BELIEF (choice, distance, token‑F1), ANSWERABILITY (list accuracy, Y/N F1, “All”), plus FACT token‑F1. Additional splits cover first‑ vs. second‑order (cyclic/acyclic) beliefs and short/full conversation contexts.

null

FANToM

Contested

Yes

No

Yes

conversation grounding (reporting bias), , adversarial false answers (wrong options are overlapped with the words in context to rule our surface matching), multi-format question sets (illusory success test by consistency checks across different formats), Manual MTurk validation (conversation-answer coherence and correctness)

Means, comparisons with percentage point gaps.

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

The reframed dataset contains questions asking how the person beliefs about certain information when the character randomly joins/leaves the conversation to create natural information gaps.

Composite phenomenon

Yes

null

No

Theory of Mind

null

['Crowd-sourced', 'LLM-generated']

['Convenience', 'Targeted', 'Criterion']

['Multiple choice', 'Short free response', 'Free response']

['Exact match', 'LLM-as-a-Judge']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean']

huInfiAgentDABenchEvaluatingAgents2024

InfiAgent-DABench: Evaluating Agents on Data Analysis Tasks

Include

null

This paper introduces InfiAgent-DABench, a benchmark designed to evaluate LLM-based agents on data analysis tasks. These tasks require agents to solve the tasks end-to-end by interacting with an execution environment.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Data Analysis

null

Data analysis is a systematic process of examining, cleaning, transforming, and modeling data to discover useful information, inform conclusions, and support decisionmaking

Subset

null

Given a file containing data and a question based on the data; generate executable code to answer the question and provide the final answer

Each item contains a question, concepts tested, constraints, answer format, file name which points to the data table and difficulty level

null

Real task examples (e.g. GitHub issues),

257

Yes

Difficulty Level, Concepts

Convenience sample (creators found a set of tasks that was readily accessible)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Another LLM is used in the pipeline to reformat the answer

Simple Mean

No

null

https://github.com/InfiAgent/InfiAgent/tree/main

huInfiAgentDABenchEvaluatingAgents2024

Widely-agreed

Yes

No

No comparisons made

The benchmark is itself realistic

No

Yes

The authors define a 6-point evaluation covering suitalbleness, reasonableness, value, restrictiveness, alignment and correctness of the generated questions and conduct a human evaluation to verify each question. 85% samples are qualified and kept in final, demonstrating the effectiveness of their dataset construction method. They also compare GPT-4 generated questions to human-made ones and show that the questions written by human experts are quite similar to those generated by GPT-4.

null

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Single cohesive phenomenon

Not applicable

null

No

Code Generation

null

['Real task', 'Unknown']

['Convenience']

['Short free response']

['Exact match']

['Widely-agreed']

['Yes']

['Realistic']

['Yes']

['Partial']

null

fanNPHardEvalDynamicBenchmark2024

NPHardEval: Dynamic Benchmark on Reasoning Ability of Large Language Models via Complexity Classes

Include

null

NPHardEval is a comprehensive benchmark consisting of 900 algorithmic tasks across P, NP-Complete, and NP-Hard complexity classes, designed specifically to evaluate large language models' reasoning capabilities. It targets to accurately assess LLMs' algorithmic problem-solving abilities across varying computational complexity levels.

(1) 9 distinct tasks × 10 difficulty levels, (2) automatic generation & verification pipeline, (3) first benchmark to ground LLM evaluation in computational‑complexity theory, (4) open‑sourced

General Capability (A broadly useful ability, which could be relevant to multiple applications)

reasoning, complexity

Yes

Reasoning is operationalised as solving decision problems in P, NP‑Complete, and NP‑Hard

Subset

null

Each task provides a specific problem scenario along with expected output format (such as True/False decisions, path lists, integers, color mappings, or other specified formats).

One problem instance + required output (True/False, path list, integer, colour mapping, etc.).

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Procedurally-generated task examples (e.g. Creating instances from a template)

900 per monthly release (100 per task)

Yes

Complexity class, task type, algorithmic difficulty level

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), LLM post-processing (extracting answers, reformatting for automated scoring)

null

Data synthesis for graph/linear tasks, difficulty depending on size/weights

Academia

Yes

MIT licensed, monthly refreshed/released the additional data (partially)

null

Test

NA

null

Weighted Mean

Yes

Task difficulty, complexity classes, model performances

null

https://github.com/casmlab/NPHardEval

NPHardEval

Contested

Yes

No

null

Wilcoxon tests, variance analysis

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

Not intended for direct end‑user interaction but aims for analytic rigor.

Composite phenomenon

Yes

null

Yes

Reasoning

null

['Author-crafted', 'Procedurally-generated']

['Targeted', 'Criterion']

['Short free response']

['Exact match', 'LLM post-processing']

['Contested']

['Yes']

['No comparison made']

['No']

['Constructed']

['Other']

wangCanLanguageModels2024

Can Language Models Serve as Text-Based World Simulators?

Include

null

This paper introduces a new benchmark, containing a dataset of text game state transitions and accompanying game tasks. They use this to directly quantify how well LLMs can serve as text-based world simulators. They test GPT-4 on this dataset and find that, despite its impressive performance, it is still an unreliable world simulator without further innovations.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

abilities of LLMs to directly simulate virtual environments

No

The phenomenon is not defined with details. I understood it as the ability to simulate the world.

Comprehensive

null

In this task, LLMs serve as world simulators in text-based virtual environments, in which an agent receives observations and proposes actions in natural language in order to complete certain objectives.

Each task is a human-authored text games that each simulate a different scientific or commonsense reasoning concept. Each item is the game state (s_t, r_t, d_t) as well as its intermediate state s_act (t+1) at each time step t as a JSON object.

null

Modified from another benchmark (e.g. translation into another language), The dataset is derived from the open BYTESIZED32 corpus.

76,369 virtual text environment state transitions

No

null

Random sample (creators defined a task space and sampled from it)

Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

No

null

https://github.com/cognitiveailab/GPT-simulator/tree/main

BYTESIZED32-State-Prediction

Not defined

Yes

No

No comparisons made

No

Yes

No

null

Simple mean

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Authors' description is unclear

Not applicable

null

No

Grounding

null

['Another benchmark', 'Another benchmark']

['Random']

['Free response']

['Exact match']

['No definition']

['Yes']

['No comparison made']

['No']

['Constructed']

['Mean']

liGSMplusComprehensiveBenchmark2024

GSM-Plus: A Comprehensive Benchmark for Evaluating the Robustness of LLMs as Mathematical Problem Solvers

Include

null

The authors create GSM‑PLUS, an adversarial extension of GSM8K that perturbs each seed question in eight different ways to test the robustness of LLM mathematical reasoning. They observe the huge performance drops from models on critical‑thinking and arithmetic variations with four prompting strategies.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

reasoning, math

Yes

Math abilities (numerical, arithmetic, understanding, distractor handling, and critical thinking) with 8 perturbation types

Subset

null

Given a word‑problem, the model must output the numerical answer.

Question context + Question (+additional constraints with variation on the question)

Robustness is measured by the consistency of solving both the seed (original) question and its perturbed variants.

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Crowd-sourced task examples (e.g. Prolific-created tasks), LLM-generated task examples (e.g. Filtered from responses to a prompt)

1319 seed (original) data, 10552 perturbed data

Yes

Seed, perturbation type, subcategory, gold answer

Targeted items (creators defined a task space and chose tasks within it strategically)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), LLM post-processing (extracting answers, reformatting for automated scoring), Distribution (perplexity, calibration, correlation)

PDR (performance drop rate), ASP (accurately-solved pairs)

GPT‑4 first rewrites each GSM8K test item, and approximately 20% of variations are manually revised by paid annotators.

Academia

Yes

https://qtli.github.io/GSM-Plus/

null

Test

null

Simple Mean

Yes

different perturbation level types (Numerical Substitution; Digit Expansion; Integer-decimal-fraction Conversion; Adding Operation; Reversing Operation; Problem Understanding; Distractor Insertion; Critical Thinking.)

null

GSM-Plus

Contested

Yes

No

Yes

Annotation consistency scores, and pass-rates to justify the validty.

Means and percentage differences

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

Yes

Reasoning

Mathematical

null

['Author-crafted', 'Crowd-sourced', 'LLM-generated']

['Targeted']

['Short free response']

['Exact match', 'LLM post-processing', 'Distribution']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean']

zhangMultiTrustComprehensiveBenchmark2024

MULTITRUST: A Comprehensive Benchmark Towards Trustworthy Multimodal Large Language Models

Include

null

The paper introduces a benchmark on the trustworthiness of MLLMs across five primarcy aspects: truthfulness, safety, robustness, fairness, and privacy. It benchmarks 20+ MLMMs and highlights the complexities introduced by multi-modality.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

trustworthiness

Yes

"Drawing on extensive studies in trustworthy LLMs and distilling from relevant literature of MLLMs, we pinpoint 5 primary aspects of trustworthiness for evaluating MLLMs, including truthfulness, safety, robustness, fairness, and privacy. In particular, truthfulness, safety, and robustness guarantee the models’ reliability and stability in preventing undesirable outcomes, i.e., errors, harms, and variations under different conditions" (page 3)

Subset

null

There are 32 tasks that are generative and/or discriminative, multimodal or test-only. Wide range from NSFW image description to PII leakage in conversations They utilize off the shelf datasets, augment existing ones, and create their own.

Varies greatly from task to task. Some include images and a prompt, others just a prompt.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language)

"more than 15k" (page 103)

Yes

image information, types of queries

Targeted items (creators defined a task space and chose tasks within it strategically), Unknown

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics), Distribution (perplexity, calibration, correlation), Correlation (Matthew's correlation, Pearson's r)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

In the main body, it's sub-aspect (e.g., truthfulness, safety, robustness). In the appendix, there are subsets of many of the tasks/existing benchmarks they run.

null

https://github.com/thu-ml/MMTrustEval

MULTITRUST

Contested

Yes

No

No comparisons made

No

null

simple mean/sum, correlation between overall rankings and general capabilities based on MMBench

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

No

Alignment

Safety

null

['Author-crafted', 'Another benchmark']

['Targeted', 'Unknown']

['Multiple choice', 'Short free response', 'Free response']

['Exact match', 'LLM-as-a-Judge', 'Distribution', 'Correlation']

['Contested']

['Yes']

['No comparison made']

['No']

['Partial', 'Representative']

['Mean', 'Other']

bittonVisITbenchDynamicBenchmark2023

VisIT-Bench: A Dynamic Benchmark for Evaluating Instruction-Following Vision-and-Language Models

Include

null

VisIT-Bench (Visual InsTruction Benchmark) is a benchmark for evaluating instruction-following vision-language models for real-world use. The authors curated 70 “instruction families” that they believe instruction-tuned vision-language models should be able to address. We conduct a large-scale empirical comparison of multimodal instruction-following models using their benchmark.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Instruction following capabilities

No

null

Subset

null

The format of the questions is either MCQ or binary QA (Yes/No). They provide a collection of 70 different open-generation tasks, like reasoning over plots, object recognition, location understanding etc.

Each instance contains an instruction, input image(s), an instruction-conditioned caption (a human-crafted caption for the image(s)/instruction), and a human-verified reference. Instructions are image-contextual imperative requests or questions, e.g., for an image of pancakes, a user asks “How can I cook this in a healthy way?”.

null

Crowd-sourced task examples (e.g. Prolific-created tasks), Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

592

Yes

instruction type, image source

Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice

n-gram (BLEU, ROUGE, chrF), Human ratings (text quality, preference, NOT manual scoring of other metrics), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics), Elo ratings, Win rate

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

stratification of results based on the instruction category

null

https://huggingface.co/datasets/mlfoundations/VisIT-Bench

VisIT-Bench

Contested

Yes

No

No comparisons made

The benchmark is itself realistic

No

null

simple mean

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

null

Composite phenomenon

Yes

null

No

Instruction Following

null

['Crowd-sourced', 'Another benchmark', 'LLM-generated']

['Targeted']

['Multiple choice']

['Soft match', 'Human ratings', 'LLM-as-a-Judge', 'Reward']

['Contested']

['Yes']

['Realistic']

['No']

['Complete']

['Mean']

singhIndicGenBenchMultilingualBenchmark2024

IndicGenBench: A Multilingual Benchmark to Evaluate Generation Capabilities of LLMs on Indic Languages

Include

null

This benchmark introduces INDICGENBENCH; a benchmark for evaluating LLMs on user-facing generation tasks across a diverse set 29 of Indic languages covering 13 scripts and 4 language families. INDICGENBENCH is composed of diverse generation tasks like cross-lingual summarization, machine translation, and cross-lingual question answering. INDICGENBENCH extends existing benchmarks to many Indic languages through human curation providing multi-way parallel evaluation data for many under-represented Indic languages.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Indic language generation capabilities

Yes

The phenomenon is defined is the ability to generate text in Indic languages.

Subset

null

Generation tasks are cross-lingual summarization, machine translation, and cross-lingual question answering.

Summarization and translation: A text sentence Question-answering: A sentence as a question and multiple choices

null

Modified from another benchmark (e.g. translation into another language)

train:~10k, test:~100k, dev:~60k

null

language

Random sample (creators defined a task space and sampled from it)

Multiple choice, Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF)

null

Industry

Yes

null

Test, Train

null

Simple Mean

Yes

language, and sub-task

null

https://github.com/google-research-datasets/indic-gen-bench/

IndicGenBench

Widely-agreed

Yes

No

No comparisons made

No

null

Simple mean

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

Yes

NLP

null

Multilinguality

['Another benchmark']

['Random']

['Multiple choice', 'Free response']

['Exact match', 'Soft match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Constructed']

['Mean']

castillo-boladoPromptsDynamicConversational2024

Beyond Prompts: Dynamic Conversational Benchmarking of Large Language Models

Include

null

This paper introduces the LTM benchmark designed to evaluate the Long-Term Memory (LTM) and Continual Learning (CL) capabilities of conversational agents. The LTM Benchmark engages agents in a single, prolonged conversation, incorporating multiple tasks and distractions to simulate realistic and meaningful interactions.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Long Term Memory

Yes

The ability to recall and effectively utilize past information. LTM encompasses several skills related to the generation and management of memories, which include but are not limited to recall, information integration, and the handling of conflicting information.

Subset

null

An LLM engages in a prolonged, dynamic conversation where multiple tasks are interleaved. Within this conversation, specific pieces of information, termed 'needles,' are included amidst unrelated content ('haystack'). These needles are essential for completing subsequent tasks. The LLM is later queried on these tasks, requiring it to retrieve and integrate the relevant needles from the conversation history, thereby assessing its long-term memory and information integration capabilities.

A dynamic, multi-turn conversation containing interleaved messages from different tasks, within which specific "needles" (relevant sentences) are included.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), LLM-generated task examples (e.g. Filtered from responses to a prompt)

33

No

null

Targeted items (creators defined a task space and chose tasks within it strategically)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF), LLM post-processing (extracting answers, reformatting for automated scoring)

null

Industry

Yes

null

Test

null

Simple Mean

Yes

They include the separate runs for each test scenario subset in the Appendix.

null

https://github.com/GoodAI/goodai-ltm-benchmark

LTM Benchmark

Widely-agreed

Yes

No

No comparisons made

No

null

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Single cohesive phenomenon

Not applicable

null

No

NLP

Long Context

null

['Author-crafted', 'LLM-generated']

['Targeted']

['Short free response']

['Exact match', 'Soft match', 'LLM post-processing']

['Widely-agreed']

['Yes']

['No comparison made']

['']

['Partial']

null

chuTimeBenchComprehensiveEvaluation2024

TimeBench: A Comprehensive Evaluation of Temporal Reasoning Abilities in Large Language Models

Include

null

TIMEBENCH is a comprehensive multi-task benchmark for evaluating large language models' temporal reasoning capabilities across symbolic, commonsense, and event-level reasoning. The benchmark evaluates various LLMs under different prompting conditions, revealing significant performance gaps compared to humans while providing detailed analyses of errors and scaling behaviors.

This work unifies previously scattered temporal reasoning datasets under a coherent taxonomic framework with standardized evaluation metrics.

General Capability (A broadly useful ability, which could be relevant to multiple applications)

temporal reasoning

Yes

TIMEBENCH focuses on a comprehensive evaluation of the temporal reasoning capabilities of large language models in challenging and complex scenarios. To achieve this goal, we summarize the difficulties and challenges faced in temporal reasoning, categorize them into three levels, and integrate diverse task formats to better align with the intricate nature of temporal reasoning (Sec 2.1)

Comprehensive

symbolic, commonsense, event relations

The model is required to use temporal reasoning to assign the correct entailment label (premise-hypothesis pair), extract the correct short answer span, and identify all correct options from a 4-choice multi-select question, or generate sentences with temporal-keywords.

Context (+ temporal keywords) + question/premise (+ answer options) to predict answers.

null

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language)

19,000

Yes

Categories (symbolic/commonsense/event), type(free-form reading comprehension, natural language inference, generation, multi-select questions), human score

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF)

null

Examples are filtered and resampled while some keeping some filtered synthetic generation.

Academia

Yes

null

Test

null

Simple Mean

Yes

Subtask, Categories

null

ERROR: type should be string, got " https://github.com/TimeBench/TimeBench"

TIMEBENCH

Widely-agreed

Yes

No

Yes

Analysis shortcomings (implicit reasoning, symbolic arithmetic)

Mean, percentage

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

3 major categories, 10 tasks, 15 subtasks within the 19,000 instances. Approximately 11k symbolic, 3.1K commonsense, 4.9k event relevant instances.

Yes

Reasoning

Temporal

null

['Real task', 'Author-crafted', 'Another benchmark']

['Convenience', 'Targeted']

['Multiple choice', 'Short free response', 'Free response', 'Structured']

['Exact match', 'Soft match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Mean']

srinivasanCLiMBContinualLearning2022

CLiMB: A Continual Learning Benchmark for Vision-and-Language Tasks

Include

null

The paper introduces CLiMB, a benchmark designed to evaluate continual learning (CL) for multimodal tasks, addressing the challenges of learning both new multimodal and unimodal tasks over time. It shows that while common CL methods can reduce forgetting in multimodal learning.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Continual learning, multimodal reasoning, knowledge transfer

Yes

Upstream Continual Learning of Multimodal Tasks: A candidate model 𝑀 M encounters a sequence of vision-language tasks... We evaluate two primary model properties in the upstream phase: upstream knowledge transfer from past learned tasks to new tasks, and withstanding forgetting of previously-seen tasks. Downstream Transfer to Low-Shot Tasks: We evaluate the low-shot adaptation ability of the model after learning each upstream vision-language task

Subset

The paper provides specific operational definitions for components like knowledge transfer, forgetting, and low-shot adaptation.

Learning from a sequence of different multimodal (vision-and-language) tasks in a continual learning (CL) setting, and then transferring to low-shot multimodal and unimodal tasks.

Each item framed as a classification problem: some input and a target label, where the inputs can be vision-only ( an image), language-only (a sentence or question), or multimodal (an image paired with a question or caption).

null

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language)

N/A

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), VQAScore

null

Academia

Yes

null

Test

null

Simple Mean

No

null

https://github.com/GLAMOR-USC/CLiMB

CLiMB

Contested

Yes

No

No comparisons made

No

null

simple mean, std, relative performance changes

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

the benchmark consists of the already existent benchmarks and doesn't provide any numbers in the paper

null

Language Modelling

Updating

null

['Real task', 'Another benchmark']

['Convenience', 'Targeted', 'Criterion']

['Short free response']

['Exact match', 'Exact match']

['Contested']

['Yes']

['No comparison made']

['No']

['Partial', 'Constructed']

['Mean', 'Std']

liVRSBenchVersatileVisionlanguage2024

VRSBench: A Versatile Vision-Language Benchmark Dataset for Remote Sensing Image Understanding

Include

null

VRSBench is a versatile vision-language dataset and benchmark for remote sensing image understanding. This comprehensive dataset not only addresses the limitations of previous datasets that either ignore detailed object information or suffer from quality control issues but also enriches the field by providing a diverse range of annotations including detailed captions, object referring, and visual question answering with rich object information and verified by human annotators.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

remote-sensing image understanding

Yes

Remote sensing models seek to understand the Earth’s surface using imagery captured from overhead, offering a unique perspective of our physical world. This technique is instrumental in various applications, such as land use mapping, urban planning, precision agriculture, disaster management, etc.

Comprehensive

null

They construct three distinct tasks for advancing remote sensing image under- standing: • VRSBench-Cap: This challenge requires the prediction of a comprehensive description for a given remote sensing image, encapsulating intricate object details and contextual relevance. • VRSBench-Ref: The task involves identifying and localizing specific objects from a given remote sensing image based on textual descriptions. • VRSBench-VQA: This task aims to answer questions related to visual content in a given remote sensing image.

Each item in the dataset consists of an image, its caption, and a list of questions and answers based on the image. Each question is accompanied by metadata indicating its type, which specifies the aspect of the image it refers to (e.g., object coordinates, object position, object category, etc.).

null

Crowd-sourced task examples (e.g. Prolific-created tasks), Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

37,408

Yes

Question type: which specifies the aspect of the image the question refers to (e.g., object coordinates, object position, object category, etc.).

Convenience sample (creators found a set of tasks that was readily accessible)

Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics), Intersection over Union (IoU)

null

Academia

Yes

null

Test, Train

Train: 85,813

null

Simple Mean

No

null

https://huggingface.co/datasets/xiang709/VRSBench

VRSBench

Contested

Yes

No

No comparisons made

No

null

simple mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

No

VQA

null

['Crowd-sourced', 'Another benchmark', 'LLM-generated']

['Convenience']

['Short free response', 'Free response']

['Exact match', 'Soft match', 'LLM-as-a-Judge', 'Soft match']

['Contested']

['Yes']

['No comparison made']

['No']

['Partial']

['Mean']

linTruthfulQAMeasuringHow2022

TruthfulQA: Measuring How Models Mimic Human Falsehoods

Include

null

Introduces TruthfulQA, a benchmark to evaluate the generation of false statements that imitate common human misconceptions across e.g. health, law, finance and politics. Suggests that scaling up models alone won't address truthfulness issues, as in their paper larger models are often less truthful.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

truthfulness

Yes

"We use a strict standard, where a claim counts as true if it describes the lit- eral truth about the real world. Claims that are only true according to a belief system or tradition (e.g. “Libras have friendly personalities”) are counted as false. TruthfulQA mostly concerns factual claims, and true factual claims are usually supported by reliable, publicly available evidence." p.3-4

Subset

null

TruthfulQA contains questions to identify answers representing common human misconceptions, requiring models to generate truthful responses instead of repeating falsehoods found in their training data.

Question along with sets of reference true and false answers, and a source.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

817

Yes

category, filtered/unfiltered

Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice, Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), Human ratings (text quality, preference, NOT manual scoring of other metrics), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

category, filtered/unfiltered

null

ERROR: type should be string, got " https://github.com/ sylinrl/TruthfulQA"

TruthfulQA

Widely-agreed

Yes

No

No comparisons made

No

Yes

They validate their reference answers through independent validators who disagreed with only 7% of the authors' judgments

Simple Means, the percentage of questions answered correctly.

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

The questions cover domains like health, law, finance, and politics where truthful AI responses would be crucial in applications.

Single cohesive phenomenon

Not applicable

817 questions: 437 "filtered" questions (adversarially filtered to be difficult for GPT-3-175B) and 380 "unfiltered" questions (expected to be difficult but not tested against the model).

No

Alignment

null

['Author-crafted']

['Targeted']

['Multiple choice', 'Short free response']

['Exact match', 'Human ratings', 'LLM-as-a-Judge']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Mean']

samdarshiConnectingDotsEvaluating2024

Connecting the Dots: Evaluating Abstract Reasoning Capabilities of LLMs Using the New York Times Connections Word Game

Include

null

The paper introduces a dataset of New York Times Connections puzzles with custom metrics to evaluate top language models against human players of varying skill levels. Results show that the best models significantly underperform both novice and expert humans. Based on the empirical analysis, the work develops a knowledge taxonomy to analyze model limitations in word categorization tasks.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

(Latent) reasoning, Lexical categorization in LLMs

Yes

Abstract reasoning represents a person’s ability to solve problems, identify patterns, and work with logical systems. We propose the NYT Connections Game as a test bed for investigating the abstract reasoning capabilities of both humans and large language models (LLMs). (Section 1)

Subset

null

Each example supplies a list of 16 unordered words taken from a single NYT Connections puzzle. The model is expected to partition them into four disjoint clusters of four words and name the underlying category for each cluster, producing its solution in one shot without feedback or retries.

A single Connections puzzle = a list of 16 words + hidden gold groupings + category labels.

null

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

438 puzzles

Yes

Category difficulty, taxonomy label for each knowledge grouping.

Convenience sample (creators found a set of tasks that was readily accessible)

Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall), Distribution (perplexity, calibration, correlation)

null

An archival site consisting of all possible answer choices and their corresponding categorizations. As the NYT does not maintain an archive of NYT Connections puzzles, we resorted to an external, third-party site for data collection. Our data spans daily problems from the conception of NYT Connections in June 2023 to August 2024. (Section 3.1.)

Academia

Yes

Human study used volunteer peers (14‑60 yrs).

null

Test

null

model outputs category name + four‑word list per line

Simple Mean

Yes

Distribution for each metric, success rates per reasoning type

null

ERROR: type should be string, got "\thttps://github.com/mustafamariam/LLM‑Connections‑Solver"

NYT Connections

Contested

Yes

Success requires broad types of knowledge

simple mean, weighted and unweighted clustering scores, frequency counts, Fleiss Kappa

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

Models/humans see identical word grids, single‑shot constraint for both.

Composite phenomenon

Yes

438 puzzles = 7,008 word instances; 1,752 category instances

Yes

Reasoning

null

['Real task', 'Author-crafted']

['Convenience']

['Structured']

['Exact match', 'Distribution']

['Contested']

['Yes']

['Comparison made']

['Yes']

['Representative']

['Mean', 'Other']

herediaXNLIeuDatasetCrosslingual2024

XNLIeu: a dataset for cross-lingual NLI in Basque

Include

null

XNLIeu is an expanded version of the XNLI benchmark that includes Basque, created by machine-translating and then manually post-editing the original English data to support cross-lingual NLI research in low-resource languages. Experiments with various LLMs show that post-editing significantly improves performance and that the translate-train strategy is most effective, though its advantage lessens when applied to natively created datasets.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Natural language understanding with the focus on cross-lingual natural language inference (NLI).

Yes

The Natural Language Inference (NLI) task consists in classifying pairs of sentences –a premise and a hypothesis– according to their semantic relation: entailment, when the meaning of the premise entails that of the hypothesis; contradiction, when both sentences have opposing truth conditions and can not co-occur at the same time; and neutral, when both sentences are not semantically related.

Subset

null

The task is to classify pairs of sentences—a premise and a hypothesis—into one of three categories based on their semantic relationship: entailment, contradiction, or neutral.

a premise, a hypothesis, and the related label (entailed, contradicts, or neutral)

null

Expert-crafted task examples (e.g. hand-written examples), Crowd-sourced task examples (e.g. Prolific-created tasks), Modified from another benchmark (e.g. translation into another language)

392, 702

No

null

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

No

null

https://github.com/hitz-zentroa/xnli-eu

XNLIeu

Contested

Yes

No

Yes

The impact of machine translation vs. professional post-edition. They justify the creation of a native Basque set to address biases and artefacts common in translation-based datasets.

simple mean and std

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

No

NLP

Understanding

Multilinguality

['Expert-crafted', 'Crowd-sourced', 'Another benchmark']

['Targeted', 'Criterion']

['Short free response']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean', 'Std']

alamCTIBenchBenchmarkEvaluating2024

CTIBench: A Benchmark for Evaluating LLMs in Cyber Threat Intelligence

Include

null

This paper introduces CTIBench, a benchmark designed to assess LLMs’ performance in cyner threat intelligence (CTI) applications. CTIBench includes multiple datasets focused on evaluating knowledge acquired by LLMs in the cyber-threat landscape.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Cyber threat intelligence

Yes

Cyber Threat Intelligence (CTI) is the ability to process and analyze vast amounts of unstructured threat and attack data; allowing security analysts to utilize more intelligence sources. It involves the collection, analysis, and dissemination of information about potential or current threats to an organization’s cyber systems can provide actionable insights to help organizations defend against these attacks.

Subset

null

The CTIBench benchmark evaluates LLMs on five cybersecurity tasks: answering multiple-choice CTI questions (CTI-MCQ), mapping vulnerabilities to their root causes (CWEs), predicting severity scores (CVSS), extracting attack techniques (MITRE ATT&CK), and attributing threats to actors or malware. Each task provides a text input (e.g. vulnerability descriptions or threat reports) and expects structured CTI-relevant outputs.

1. CTI-MCQ: One row contains a multiple-choice question with a question string, four answer options, and the correct answer label. 2. CTI-RCM: A CVE description (free-text) and the corresponding CWE label representing the root cause. 3. CTI-VSP: One row has a CVE description and the associated CVSS v3.1 vector string with detailed severity metrics. 4. CTI-ATE: One row contains a description of a threat behaviour and a list of MITRE ATT&CK technique IDs mentioned in the report. 5. CTI-TAA: Each row contains a threat report from a reputed vendor mapped to an Avanced Persistent Threat (APT) group.

null

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), LLM-generated task examples (e.g. Filtered from responses to a prompt)

4947

No

null

Random sample (creators defined a task space and sampled from it), Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice, Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

Scores are presented separately for each task subset

null

https://github.com/aiforsec/cti-bench

CTIBench

Widely-agreed

Somewhat. Certain tasks in the benchmark align well with how real-world analysts evaluate cyber threat intelligence, suggesting some face validity. However, other tasks focus more on knowledge retrieval, which may not reflect the full nature of cyber threat intelligence, where knowledge retrieval, understanding, reasoning, and application are all important. These aspects are tested separately, so the benchmark doesn’t provide a full picture of end-to-end evaluation.

Yes

No

No comparisons made

No

The authors sample task items from real-world cyber threat issues and reformat them for evaluation. While this means the tasks are grounded in real-world problems, it remains unclear whether the evaluation of LLMs aligns with how cyber threat analysts would perform such assessments.

null

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

No

Alignment

Safety

null

['Real task', 'Author-crafted', 'LLM-generated']

['Random', 'Convenience', 'Targeted']

['Multiple choice', 'Short free response']

['Exact match']

['Widely-agreed']

['Partially']

['Yes']

['No comparison made']

['No']

['Partial', 'Representative']

null

jinJailbreakingLargeLanguage2024

Jailbreaking Large Language Models Against Moderation Guardrails via Cipher Characters

Include

null

The paper proposes JAMBench - a benchmark for LLM jailbreaks against content moderation classifiers used as guardrails. The dataset is specifically designed to not trigger input-level harm classifiers, but trigger output-level harm classifiers to enable the study of evading output-level detection through jailbreaks.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

natural language understanding, jailbreaking, harmful behaviour

No

Vulnerability of content moderation classifiers to jailbreaks.

Subset

null

The task is to elicit a harmful response by an LLM while circumventing content moderation.

A harmful question

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

160

Yes

Category, severity of harm

Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice

Exact Match (accuracy, F1, precision, recall), A non-defined "jailbreak success rate". likely LLM-as-a-Judge but unclear.

null

author-crafted from scratch.

Academia

Yes

null

Extremely little information is provided about the process and rationale of creating items. However, it is clear that datasets are designed to be adversarial to current content moderation. This has major implications on the construct validity but is not discussed at all.

Test

null

The base task is eliciting harmful response from an LLM while evading content moderation (so the LLM outputs Free Response). However, for the benchmark, we are utilizing [Harmful Question + LLM + Content Moderator ] and receive a harmfulness score that should be deceptive and hence the wrong filter decision is made.

Simple Mean

Yes

harm domain: Hate, Sexual, Violence, Self-Harm

null

https://github.com/Allen-piexl/llm_moderation_attack

JAMBench

Not defined

Yes

No

null

Experiments are repeated 5 times but resulting information onf uncertainty are not reported.

Model access required (e.g. logits)

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

Yes

Alignment

Safety

null

['Author-crafted']

['Targeted']

['Multiple choice']

['Exact match', 'Reward']

['No definition']

['Yes']

['No comparison made']

['No']

['Partial']

['Unknown']

zhouVLUEMultitaskMultidimension2022

VLUE: A Multi-Task Multi-Dimension Benchmark for Evaluating Vision-Language Pre-training

Include

null

VLUE is a vision-language benchmark consisting of 4 representative VL tasks, each equipped with a private test set annotated on images from wild distribution. The authors evaluate the efficiency-performance trade-off of representative VLP models and build a Pareto SOTA landscape for current VLP research. Additionally they provide an extensive analysis of the generalization ability and the efficiency-performance trade-off of representative VLP models.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

- Generalization and transferability of VLP models - Efficiency-performance trade-off

No

The authors do not explicitly define the above phenomena. They only mention the following: - Generalization: VLP models have already seen the images in the downstream datasets and their captions before fine-tuning and evaluating on them, overestimating their transfer and generalization abilities. - Performance-efficiency: "We refer to the goal of this phenomenon as “Pareto SOTA” following, which means that there is no other model currently better than it on all the dimensions of interest such as performance and efficiency. Therefore, we believe it is necessary to measure and report performance-efficiency trade-off."

Comprehensive

null

VLUE covers a set of fundamental VL tasks, including image-text retrieval, visual question answering, visual reasoning, and visual grounding.

Depending on the task: - retrieval: query and the relevant image - reasoning: pair of images and a natural statement, True/False label - grounding: image, target object, target location - QA: image, question and answer - captioning: image and caption

null

Crowd-sourced task examples (e.g. Prolific-created tasks), Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

74,509

No

null

Convenience sample (creators found a set of tasks that was readily accessible)

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train

more than 2M

null

Simple Mean

No

null

https://github.com/MichaelZhouwang/VLUE/tree/main

VLUE

Not defined

No

Yes

No

No comparisons made

No

null

simple mean

Outputs alone

Representative task (e.g. answering medical licensing exam questions), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

No

Language Modelling

Adaptability

null

['Crowd-sourced', 'Another benchmark', 'LLM-generated']

['Convenience']

['Multiple choice', 'Short free response', 'Free response']

['Exact match', 'Soft match']

['No definition']

['No']

['Yes']

['No comparison made']

['No']

['Representative', 'Constructed']

['Mean']

liuExposingAttentionGlitches2023

Exposing Attention Glitches with Flip-Flop Language Modeling

Include

null

This paper introduces FFLM, a synthetic benchmark designed to investigate "attention glitches" that cause Transformer-based language models to make sporadic reasoning errors.The benchmark tests models' ability to copy binary symbols across long distances while ignoring intervening tokens. The paper with this benchmark further explores regularisation and architectural tweaks to mitigate them.

Provides large‑scale empirical study (10 k+ models) and mechanistic analysis.

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Reasoning, (Robust long‑range) memory

Yes

A flip-flop language modeling (FFLM) task is defined on sequences of write, read, and ignore instructions: write sets the memory state to a certain value which is later retrieved by read, while ignoring any contents in between. Authors are interested in whether language models can learn a flip-flop language from samples, which they define as processing the read operations perfectly to understand model output inaccuracies. (Sections 1 and 3.1)

Subset

They assume that the correct memory retrieval (subset) as the absence of incorrect responses.

Predict the next token in sequences built from write/read/ignore instructions plus binary data so that every read must output the latest write bit.

A length‑T token sequence, evaluation checks all read positions.

null

Procedurally-generated task examples (e.g. Creating instances from a template)

16M training sequences

Yes

difficulty (fixed sequence lengths, writing probability, reading probability, computationally ignoring probability)

Random sample (creators defined a task space and sampled from it)

Short free response (e.g. single word or number), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train, Validation

null

Deterministic read position index, binary digits

Simple Mean

Yes

Three canonical splits: (1) In-distribution, (2) sparse out-of-distribution token sets, (3) dense out-of-distribution sets where tokens are frequently appearing in read/write functions

null

https://huggingface.co/datasets/synthseq/flipflop

FFLM

Contested

Yes

No

Yes

500 seed replicate study

Mean error, scatter plots, attention heatmaps

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

Synthetic stress‑test (not user interaction scenarios)

Single cohesive phenomenon

No

16M training sequences, 160K sparse o.o.d. sequences , and 4K dense o.o.d. sequences

Yes

NLP

Long Context

null

['Procedurally-generated']

['Random']

['Short free response', 'Structured']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean']

merdjanovskaNoiseBenchBenchmarkingImpact2024

NOISEBENCH: Benchmarking the Impact of Real Label Noise on Named Entity Recognition

Include

null

NOISEBENCH evaluates label noise in Named Entity Recognition (NER) models. It provides multiple variants of the same dataset with different types of real noise (expert errors, crowd- sourcing errors, automatic annotation errors and LLM errors).

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Label Noise on Named Entity Recognition

Yes

"Benchmark for measuring the impact of label noise in the training data on the prediction quality of trained NER models"

Subset

null

Named entity recognition (NER), which requires detecting and classifying named entity types in text.

Sentence where named entities must be identified and classified into one of four entity types (PER, ORG, LOC, MISC).

null

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language)

Test: 3,427 sentences (test set)

Yes

noise level percentage, error types, entity counts

Specific criteria (items were taken from a larger set based on specified rules)

Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Benchmark is derived from the CoNLL-03 dataset, which consists of real news articles with manually annotated named entities

Academia

Yes

null

Test, Train, Validation

Train: 5,885 sentences; Validation: Approximately 17%

null

Simple Mean

Yes

Noise types (Expert, Crowd++, Crowd, Distant, Weak, LLM); Per-class metrics (LOC, ORG, PER, MISC); Token-level vs. entity-level F1 scores

null

https://github.com/elenamer/NoiseBench

NOISEBENCH

Widely-agreed

Yes

No

Yes

When comparing real noise to simulated noise, they provide evidence showing how models immediately memorize real noise patterns rather than going through distinct learning phases.

Micro-averaged entity-level F1 score reported as means across 3 runs with standard deviations. Simple means used for comparing approaches across different noise types.

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

Named entity recognition is a common NLP task with potential real-world applications.

Composite phenomenon

Yes

test set: 3,427 sentences, 5,725 entity mentions; training set: 5,885 sentences, 9,685 entity mentions.

No

NLP

Extraction

null

['Real task', 'Another benchmark']

['Criterion']

['Structured']

['Exact match']

['Widely-agreed']

['Yes']

['Comparison made']

['Yes']

['Representative']

['Mean', 'Std']

jinMMToMQAMultimodalTheory2024

MMToM-QA: Multimodal Theory of Mind Question Answering

Include

null

The paper introduces MMToM-QA, the first benchmark that evaluates Theory-of-Mind (ToM) reasoning across multimodal inputs (video and text), containing diverse test questions and synthetic training videos. The authors propose BIP-ALM, a novel approach that combines Bayesian inverse planning with language models to extract unified representations from multimodal data, demonstrating that while current large language and multimodal models lack robust ToM capacity, their method narrows the gap to human-level performance.

New multimodal Theory-of-Mind benchmark (text + video)

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Theory-of-Mind, (social) reasoning, multi-modal

Yes

While the recent ToM benchmarks provide welldesigned, cognitively informed tools, they share several notable limitations. One such limitation is the dependence on massive training data, which raises the concern that these models work by finding data patterns in a way that deviates from humanlike ToM reasoning. But ToM reasoning goes beyond merely text comprehension or video understanding. Hence the benchmark targets to evaluate how a model can infer mental states from either words or vision separately or fuse the separate information to form a single coherent mental scene.

null

For each item, the model receives a short household video clip and a textual description of the scene and actions, then asked to decide which of two candidate mental‑state hypotheses is more plausible.

One clip (RGB‑D frames) + accompanying textual scene/action description + 1 question with two options.

null

Procedurally-generated task examples (e.g. Creating instances from a template), LLM-generated task examples (e.g. Filtered from responses to a prompt)

1000 videos

Yes

Question type, timestamps, ground‑truth goals/beliefs, scene graphs, RGB‑D, segmentation, 3‑D poses, camera parameters

Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

binary options (overall chance = 50%)

Dataset items are entirely synthetic: avatar videos rendered in simulator; scene & action narratives drafted via templates and polished by GPT‑4.

Academia

Yes

null

Test, Train

null

Simple Mean

Yes

True/False belief, short/long-term, Goal inference given true/false/updated beliefs or given future actions

null

https://chuanyangjin.com/mmtom-qa

MMToM‑QA

Contested

Yes

No

No comparisons made

No

Yes

Qualitative analysis and ablations to show why their method (BIP-ALM) succeeds where LLMs fail.

simple mean, statistical tests

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

High controllability, annotation (but lacks real‑world sensor noise), human test set for generalisation.

Composite phenomenon

Yes

Test (600 Questions over 134 videos), Validation (not separately mentioned)

No

Theory of Mind

null

['Procedurally-generated', 'LLM-generated']

['Targeted']

['Multiple choice']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean', 'Tests']

dengCOLDBenchmarkChinese2022

COLD: A Benchmark for Chinese Offensive Language Detection

Include

null

COLDATASET is dataset of Chinese sentences with binary offensive/non-offensive labels covering topics of race, gender, and region. COLDETECTOR is a baseline detector trained on this dataset.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Offensive language detection in Chinese

Yes

"The contents with any form of targeted offense to individuals or groups are consid- ered offensive language. It includes veiled or di- rect offensive content expressing rudeness, disrespect, insults, threats and profanity based on as- pects such as race, religion, sex, or sexual orientation"

Subset

null

The task involves classifying Chinese text samples as either "offensive" or "non-offensive," where offensive/toxic language, and hate speech are not distinguished.

Chinese Social Media text sample with a binary label (offensive/non-offensive), and in the test set, a more fine-grained label classifying it into one of four subcategories.

null

Real task examples (e.g. GitHub issues)

test: 5,323

Yes

topic, fine-grained labels, average character length, topic-related keywords presence

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train

train/dev: 32,157

null

Simple Mean

Yes

Four subcategories (attacking individuals, attacking groups, anti-bias, other non-offensive) and also by topic categories.

null

https://github. com/thu- coai/COLDataset

COLD

Contested

Yes

No

Yes

First, report high inter-annotator agreement. Second, compare against the translated Jigsaw dataset. Third, conduct an ablation study.

Reporting accuracy, precision, recall, and F1 scores, both as macro averages across all categories and separately for offensive and non-offensive classes.

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

The classification of content as offensive/non-offensive directly mirrors a real-world content moderation task performed on social media platforms. However, in production systems, this task would likely include additional context (like user history, reports, engagement metrics).

Composite phenomenon

Yes

null

No

Alignment

Safety

null

['Real task']

['Targeted', 'Criterion']

['Multiple choice']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Partial']

['Mean']

krojerAreDiffusionModels2023

Are Diffusion Models Vision-And-Language Reasoners?

Include

null

GDBench is a benchmark designed to assess vision-and-language reasoning in diffusion-based models using image-text matching. GDBench aggregates 8 existing datasets/benchmarks to measure text retrieval, image retrieval, and bias towards religious groups, national identity, and sexual orientation. The code and benchmark setup are publicly available.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

compositional reasoning, image-text matching, vision-and-language reasoning

No

Vision-and-language reasoning is implicitly defined as the "fine-tuned understanding of vision and language," and the ability for models "to understand how various objects and parts compose together" (1).

Subset

null

"We present our method Diffusion Image-Text Matching (ITM). Our goal is to assign a score to an image(x)-text(w) pair (x, w) which is broadly useful for downstream applications. We provide (x, w) to the diffusion model and task it to “edit” the image according to the text. Our main intuition is if the image is not described by the text, a lot of edits are needed to fit the text, in which case it gets a low score, and vice-versa" (4). The benchmark includes "7 ability-centric" ITM tasks, and one bias task (6).

A single item would be an image, and a text caption, which can either be a hard negative or a positive caption.

Broadly, the tasks of GDBench are split into image retrieval tasks and text retrieval tasks. The benchmark is designed primarily for diffusion models.

Modified from another benchmark (e.g. translation into another language)

null

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Specific criteria (items were taken from a larger set based on specified rules)

Generated image

Distribution (perplexity, calibration, correlation)

The paper defines a custom metric for the normalized image retrieval error on page 5. The metric is intended to measure the "relative difference of how much easier or harder it becomes to denoise the image with a given text relative to when no text is given" (5). The metric is intended for diffusion models. The paper also measures the religious, nationality, and sexual orientation biases in the image outputs using effect size, or the defined "normalized association score" on page 7.

The 8 distinct tasks are based on 8 existing datasets (Flickr30k, Winoground, Aro, ImageCoDe, SVO, CLEVR, Pets) instead of the creation of an entirely new dataset. The GDBench then reports scores on the sub-elements per sub-part of the subsumed benchmarks.

Mix (multiple authors from industry and academia)

Yes

null

Though the benchmark aggregates existing datasets, it is included due to its contribution in creating a more extensive and comprehensive benchmark and its extension of image-text matching tasks to diffusion models.

Test

null

The model must edit the provided image until it matches the provided positive or negative caption.

null

Yes

SVO: Verb, Subj, Obj; ImageCoDe: Static, Video; ARO: VG Attr., VG Rel., COCO Ord., Flickr Ord. Bias: Religion, Nationality, Sexuality

null

https://github.com/McGill-NLP/diffusion-itm

GDBench (Generative-Discriminative Evaluation Benchmark)

Contested

Yes

No

No comparisons made

No

Yes

The authors highlight that image-text-matching (ITM) tasks have "become a standard paradigm for diagnostic vision-and-language datasets" because they enable "interpretable evaluation on many downstream skills" (6). Thus, the authors implicitly position the dataset as a valid construct to measure vision-and-language reasoning.

Image retrieval error, effect score bias

Model access required (e.g. logits)

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

No

Grounding

null

['Another benchmark']

['Convenience', 'Criterion']

['Free response']

['Distribution']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean']

tsurutaSARSCoV2InteractionDataset2024

A SARS-CoV-2 Interaction Dataset and VHH Sequence Corpus for Antibody Language Models

Include

null

AVIDa-SARS-CoV-2 is a dataset featuring the antigen-variable domain of heavy chain of heavy chain antibody (VHH) interactions obtained from two alpacas immunized with severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) spike proteins. AVIDa-SARS-CoV-2 includes binary labels indicating the binding or non-binding of diverse VHH sequences to 12 SARS-CoV-2 mutants, such as the Delta and Omicron variants. The authors report benchmark results for predicting SARS-CoV-2-VHH binding using VHHBERT pre-trained on VHHCorpus-2M and existing general protein and antibody-specific pre-trained language models.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

binding prediction, i.e. predict the binding site of an antibody.

Yes

The antibody discovery task is a binary sequence classification that distinguishes antibodies that bind to SARS-CoV-2.

Subset

null

Binary classification: whether the antibody binds to a specific antigen at the antibody sequence level.

- VHH_sequence: Amino acid sequence of VHH - Ag_label: Antigen Type - label: Binary label represented by 1 for the binding pair and 0 for the non-binding pair - subject_species: Species of the subject from which VHH was collected - subject_name: Name of the subject from which VHH was collected - subject_sex: Sex of the subject from which VHH was collected

null

Real task examples (e.g. GitHub issues)

77,003

No

null

Convenience sample (creators found a set of tasks that was readily accessible)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Industry

Yes

null

Test, Train

2M

null

Simple Mean

No

null

https://huggingface.co/datasets/COGNANO/AVIDa-SARS-CoV-2

AVIDa-SARS-CoV-2

Widely-agreed

Yes

No

No comparisons made

The benchmark is itself realistic

No

null

simple mean

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

null

Single cohesive phenomenon

Not applicable

null

No

Biology

null

['Real task']

['Convenience']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['Realistic']

['No']

['Complete']

['Mean']

maAgentBoardAnalyticalEvaluation2024

AgentBoard: An Analytical Evaluation Board of Multi-turn LLM Agents

Include

null

This paper introduces AGENTBOARD, a benchmark designed to evaluate LLM agents capabilities in partially observable environments, multi-round interactions and diverse tasks through a unified evaluation framework and fine-grained metric analysis.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Agent Abilities for Real-World Problem Solving

Yes

The ability to perform step-by-step planning in diverse tasks and partially observable environments via multi-turn interactions.

Comprehensive

null

A task in AGENTBOARD presents an agent with a real-world scenario—such as an embodied, game, or tool-use environment—where it must perform actions, receive feedback, and plan over multiple rounds in a partially observable setting.

A task item includes an environment definition, a goal to achieve, step-wise observations, and defined action space

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language)

1013

Yes

Difficulty

Convenience sample (creators found a set of tasks that was readily accessible)

Short free response (e.g. single word or number), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

Across each subtask dataset, difficulty and subskill

null

https://huggingface.co/datasets/hkust-nlp/agentboard/tree/main

AGENTBOARD

Contested

No

Yes

No

No comparisons made

The benchmark is itself realistic

No

null

Simple mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

No

Reasoning

Planning

null

['Author-crafted', 'Another benchmark']

['Convenience']

['Short free response', 'Structured']

['Exact match']

['Contested']

['No']

['Yes']

['Realistic']

['No']

['Partial']

['Mean']

maruNibblingHardCore2022

Nibbling at the Hard Core of Word Sense Disambiguation

Include

null

The authors introduce new challenging test sets for Word Sense Disambiguation evaluation specifically designed to evaluate model resilience on rare word senses and present a more rigorous evaluation framework.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Word Sense Disambiguation (WSD): automatically assigning a correct meaning to an ambiguous word in context

Yes

"Word Sense Disambiguation (WSD), the task of automatically assigning a meaning to an ambiguous word in context"

Subset

null

Word Sense Disambiguation (WSD) is the task of automatically assigning the correct meaning to an ambiguous word within a given context, selecting from a predefined sense inventory.

A single item consists of a context (sentence or paragraph) containing a target ambiguous word.

null

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language)

7,253 instances

Yes

word sense frequency, domain information, presence in training data

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

micro-averaged F1 and macro-averaged F1 scores

Yes

Presence in training data (SemCor; non-SemCor instances); Word sense frequency (WordNet first sense vs non-first sense); Dataset (ALL, ALLNEW, S10NEW, 42D, hardEN, softEN)

null

https://github.com/SapienzaNLP/wsd-hard-benchmark

null

Widely-agreed

Yes

No

Yes

null

F1 scores (micro-averaged and macro-averaged) as the primary statistical method.

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

WSD is a fundamental NLP capability that would be used as a component within larger systems.

Single cohesive phenomenon

Not applicable

null

No

NLP

Understanding

null

['Real task', 'Another benchmark']

['Targeted', 'Criterion']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Mean']

huangConMeRethinkingEvaluation2024

ConMe: Rethinking Evaluation of Compositional Reasoning for Modern VLMs

Include

null

ConMe is a multimodal compositional reasoning benchmark that presents a novel automatic hard negative generation pipeline using VLMs. It is publicly-available, manually verification, and presents a complementary automatic analysis and error categorization pipeline.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

compositional reasoning

Yes

Compositional reasoning "is the ability of the VLM to recognize and attend to the language concepts beyond objects (i.e., nouns), such as attributes, relations, finegrained object alternatives, and more, in both the image and text of a VL pair" (1).

Subset

null

A model is given an image, a correct caption, and a hard negative caption, and must choose the correct caption.

A single item contains the image, the generated correct caption, and the generated hard negative caption.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

24347

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice

Exact Match (accuracy, F1, precision, recall), Distribution (perplexity, calibration, correlation)

The benchmark evaluates sample accuracy and perplexity.

ConMe modifies the SugarCrepe examples with their own novel text generation pipeline for VLM-created hard negatives. First, GPT-4V generates "a detailed description of the input image," used as the "ground truth" (4). Then, 4 downstream VLMs (LLaVA 1.6-7b, LLaVA 1.5-7b, InstructBLIP Flan-T5, InstructBLIP Vicuna-7b) then generate descriptions of the image. Next, GPT-4V receives the description it produced, and the descriptions of the VLMs, and is prompted "to generate multiple challenging compositional reasoning questions based on the generated descriptions" from the VLMS (4). The models then cycle through disregarding the GPT-4V generated samples that all VLM models correctly answer, and improving the kept samples to become more challenging. The authors manually verified a random subsample of 1000 samples from ConMe.

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

No

null

https://github.com/jmiemirza/ConMe

ConMe (Confuse Me)

Contested

Yes

No

Yes

The authors highlight that image-to-text matching tasks rely on "LLM only negative text generation pipeline(s)" that produces "improbable" or "outlier" captions for the given image (1). As a result, the authors claim that using VLMs in the hard negative generation pipeline is required for a more accurate benchmark of compositional reasoning in VLMS.

Simple mean/accuracy

Model access required (e.g. logits)

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

No

ConMe uses three partitions used in SugarCrepe: replace-att, replace-obj, and replace-rel hard negatives. ConMe uses the same 3846 images as SugarCrepe, but generates more examples per image across the three partitions.

No

Reasoning

Compositional

null

['Author-crafted', 'Another benchmark', 'LLM-generated']

['Convenience', 'Targeted']

['Multiple choice']

['Exact match', 'Distribution']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean']

chenCurriculumBroadcoverageBenchmark2022

Curriculum: A Broad-Coverage Benchmark for Linguistic Phenomena in Natural Language Understanding

Include

null

Current models do not provide insight into how well a language model captures distinct linguistic skills essential for language understanding and reasoning. In this paper, authors introduce CURRICULUM as a new format of NLI benchmark for evaluation of broad-coverage linguistic phenomena. CURRICULUM contains a collection of datasets that covers 36 types of major linguistic phenomena and an evaluation procedure for diagnosing how well a language model captures reasoning skills for distinct types of linguistic phenomena.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Linguistic phenomena in NLU

Yes

For this phenomenon, authors try to measure how well a language model captures distinct linguistic skills essential to language understanding and reasoning.

Comprehensive

null

Natural language inference (NLI). More specifically, authors provide a group of tasks motivated by three benchmarks: GLUE Diagnostic, Rainbow, and DNC. In addition, we include many more subtasks focusing on complex reasoning types such as deductive logic and analytical thinking.

Each single item has a premise, a hypothesis, and a target label.

null

Modified from another benchmark (e.g. translation into another language)

171,252

Yes

difficulty level

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), Correlation (Matthew's correlation, Pearson's r)

null

Academia

Yes

null

Test, Train

540,437

null

Simple Mean

Yes

sub-phenomenon, difficulty

null

https://github.com/eric11eca/curriculum-ling?tab=readme-ov-file

CURRICULUM

Contested

Yes

No

null

simple mean

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

No

NLP

Understanding

null

['Another benchmark']

['Convenience', 'Targeted']

['Short free response']

['Exact match', 'Correlation']

['Contested']

['Yes']

['No comparison made']

['No']

['Constructed']

['Mean']

xieOSWorldBenchmarkingMultimodal2024

OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments

Include

null

OSWorld introduces a scalable, executable computer environment supporting real OSs (Ubuntu, Windows, macOS) to evaluate multimodal agents on 369 open-ended real-world tasks. It includes complex setups, execution-based evaluation, and detailed analysis of LLM/VLM agents' capabilities and deficiencies.

They put significant emphasis on that they are "The first-of-its-kind scalable, real computer environment for multimodal agents, supporting task setup, execution-based evaluation, and interactive learning across operating systems." <- with emphasis on the first

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Multimodal tool use, open-ended task execution in real OS environments, reasoning (though they don't mention word reasoning even once, the tasks implicitly require reasoning)

No

The phenomenon is implicitly defined through the benchmark’s design and the types of tasks it includes. Authors describe the benchmark as evaluating agents’ ability to complete open-ended, real-world computer tasks using multimodal perception and actions (such as screenshots, accessibility trees, mouse/keyboard inputs) across various applications and operating systems.

Subset

null

Open-ended computer activity, which is described in natural language, executed by an agent inside a real operating system environment. Each task includes an initial state setup, a goal instruction, and a custom execution-based evaluation script to determine success.

"Each example is carefully annotated with a natural language instruction, a setup configuration with corresponding files and setup actions for initialization of initial states upon our provided VM image, and a manually crafted evaluation script to check if the task is successfully executed." (page 7)

null

Human exam questions (e.g. GRE questions), Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Expert-crafted task examples (e.g. hand-written examples), Modified from another benchmark (e.g. translation into another language)

412

Yes

human difficulty, task feasibility, application domain, task type

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code), Extended interaction (e.g. conversation, calling an API and processing the response)

Exact Match (accuracy, F1, precision, recall), Execution-based evaluation scripts

null

They sourced tasks from an incredible number of sources, diverse between each other - there is a table (appendix B.3) taking almost two pages to just list program sourced.

Academia

Yes

null

Test

null

Simple Mean

Yes

By application domain: OS, Office, Daily, Professional, Workflow By OS: Ubuntu or Window By task difficulty: Easy, Medium, Hard (based on human completion time) By feasibility: Feasible and Infeasible tasks By input modality: Screenshot, Accessibility Tree, SoM etc.

null

https://os-world.github.io/

OSWorld

Not defined

Yes

No

Yes

null

1. Human Baseline They conducted human evaluations across all tasks, showing that humans (without prior expousure to the tasks) achieved +/- 72.36% accuracy, while top models performed under 12.24%, showing that the tasks are achievable. 2. Realistic Task Design The tasks are based on real-world scenarios, sourced from impressively large number of sources: from user forums, tutorials, and many others, to everyday computer workflows. 3. Execution-Based Evaluation They designed 134 custom, deterministic evaluation scripts to assess functional correctness and objective and reproducible scoring. 4. Model Performance Analysis The authors analysed how models fail - such as difficulty with GUI grounding, interaction noise, and poor generalisation across applications - they tried aligning that observed performance with the skills they intended to measure through the benchmark. 4. Comparative Difficulty They compare OSWorld to other benchmarks like WebArena and show that OSWorld tasks take longer for humans to complete and are harder for models to solve to support the idea that OSWorld includes more complex tasks, which are supposed to be closer to the real-world abilities.

Just simple mean, with occasional reporting of variance or distribution plots.

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

The benchmark involves full workflows from A to Z - such as software installation, document editing, web navigation, and multi-application coordination. These tasks are executed in real OS environments (Ubuntu and Windows) and they use apps/programs (e.g. Chrome, LibreOffice, VLC), to operate as general-purpose assistants.

Composite phenomenon

Yes

"OSWORLD benchmark [...] encompasses 369 real computing tasks defined and executed on Ubuntu. Additionally, we provide a set of 43 tasks for Windows built on the OSWORLD environment." (page 6)

Yes

Agents

Tool Use

null

['Human exams', 'Real task', 'Author-crafted', 'Expert-crafted', 'Another benchmark']

['Targeted', 'Criterion']

['Multiple choice', 'Short free response', 'Free response', 'Interaction']

['Exact match', 'Reward']

['No definition']

['Yes']

['No comparison made']

['']

['Complete']

['Mean', 'Std']

dasEXAMSVMultidisciplineMultilingual2024

EXAMS-V: A Multi-Discipline Multilingual Multimodal Exam Benchmark for Evaluating Vision Language Models

Include

null

EXAMS-V is a multi-discipline multimodal multilingual exam benchmark for evaluating vision language models. The questions come in 11 languages across 20 school disciplines. The evaluation results demonstrate that this is a challenging dataset, which is difficult even for advanced vision–text models such as GPT-4V and Gemini.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Multilingual and multilingual model’s multitask accuracy (knowledge acquired during pretraining) across a diverse set of subjects

No

For LLM evaluation, standardized testing akin to school examinations has proven to be an effective measure of a model’s capabilities.

Comprehensive

null

The task is a multiple-choice question answering task.

language, subject, grade, question, choices, answers, image, image type

null

Human exam questions (e.g. GRE questions), Modified from another benchmark (e.g. translation into another language)

4,800

Yes

language, grade, subject

Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test, Train

16,500

null

Simple Mean

Yes

subject, language

null

https://huggingface.co/datasets/Rocktim/EXAMS-V

EXAMS-V

Widely-agreed

Yes

No

No comparisons made

The benchmark is itself realistic

No

null

simple mean

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

No

VQA

null

['Human exams', 'Another benchmark']

['Criterion']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['Realistic']

['No']

['Representative']

['Mean']

jiangBRAINTEASERLateralThinking2023

BRAINTEASER: Lateral Thinking Puzzles for Large Language Models

Include

null

The paper introduces BRAINTEASER, a multiple‑choice benchmark that probes large language models’ ability for lateral thinking, a creative, non‑linear reasoning that overrides default commonsense associations. It describes a three‑step construction pipeline (web crawling and filtering of puzzles, semi‑automatic distractor generation, and semantic + contextual reconstructions) that yields high‑quality items while controlling for memorization.

First publicly-‑available benchmark for lateral‑thinking evaluation, Detailed error analysis regarding memorisation & commonsense traps.

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

(Lateral/creativee) reasoning

Yes

The paper targets lateral reasoning, defined as a creative problem-solving approach that differs from vertical thinking (commonsense association and inference). It requires solving puzzles that cannot be resolved through straightforward commonsense associations alone, demanding non-linear reasoning patterns.

Subset

- Although authors note lateral thinking comprises four skills (awareness, random stimulation, alternatives, alteration), they do not operationalize them separately; sub‑elements are not separately measured. - Authors explicitly contrast lateral vs vertical thinking and position benchmark as complementary to commonsense QA suites.

Multiple-choice QA where the model selects the correct explanation to a brain‑teaser puzzle among 4 options (one may be “None of the above”).

One puzzle = question stem + 4 answer choices

Sentence/word puzzles -- each puzzle has 2 reconstruction variants (semantic/context) to resist memorization.

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), LLM-generated task examples (e.g. Filtered from responses to a prompt)

1,119 puzzles

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

Group accuracy = 1 only if model answers all three variants correctly.

373 core puzzles crawled from riddles.com, rd.com, etc., distractors via COMET (commonsense pre-trained language models), reconstruction prompts via GPT‑4

Mix (multiple authors from industry and academia)

Yes

null

Test

Core 373 originals expanded 3 times, with reconstructions.

null

Simple Mean

Yes

Separate scores for Sentence/ Word and for Original/Semantic/Context splits.

null

https://github.com/1171-jpg/BrainTeaser

BRAINTEASER

Contested

Yes

No

Yes

- The human annotators rated 99% of the original QA pairs valid, and 97%-100% of the semantic/context reconstructions as consistent with the original QA pairs.

Simple proportion, human/model comparisons

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

The puzzles are artificial and deliberately engineered, so they are unlikely to appear in day‑to‑day user interactions. Rather, they approximate the kind of creative‑reasoning challenges that could arise across diverse downstream tasks. Thus ecological validity seems to be low, serving research evaluation rather than direct operational use.

Single cohesive phenomenon

Not applicable

627 sentence, 492 word; originals + reconstructions

Yes

Reasoning

null

['Real task', 'Author-crafted', 'LLM-generated']

['Convenience', 'Targeted']

['Multiple choice']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Complete', 'Constructed']

['Mean']

ushioGenerativeLanguageModels2022

Generative Language Models for Paragraph-Level Question Generation

Include

null

QG-Bench, a comprehensive benchmark for paragraph-level question generation (QG) that unifies existing question answering datasets into a standard format. The authors fine-tune LMs for the QG task and evaluate them using both automatic metrics and human evaluation.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Question generation

Yes

"Question generation is the task of generating a question given an in- put context consisting of a document, a paragraph or a sentence, and an answer where the question is anchored"

Subset

null

Generate a natural language question given an input paragraph and an answer span that appears within that paragraph.

Paragraph, a sentence within that paragraph, an answer span, and the target question to be generated.

null

Modified from another benchmark (e.g. translation into another language)

SQuaD train: 75,722

Yes

language, domain, average patagraph character length

Convenience sample (creators found a set of tasks that was readily accessible), Specific criteria (items were taken from a larger set based on specified rules)

Free response (e.g. summary paragraph, executable code)

n-gram (BLEU, ROUGE, chrF), Human ratings (text quality, preference, NOT manual scoring of other metrics), Correlation (Matthew's correlation, Pearson's r)

null

Academia

Yes

null

Test, Train, Validation

SQuaD validation: 10,570, SQuaD test: 11,877

null

Simple Mean

Yes

Scores by language, domain, model input type

null

https://github.com/asahi417/lm-question-generation

QG-Bench

Widely-agreed

Yes

No

Yes

Validation of the benchmark through manual evaluation where human annotators rate generated questions across three criteria (grammaticality, understandability, and answerability).

Simple mean scores for each metric. For correlation analysis between automatic metrics and human judgments: Spearman's rank correlation coefficient.

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

No

NLP

Extraction

null

['Another benchmark']

['Convenience', 'Criterion']

['Free response']

['Soft match', 'Human ratings', 'Correlation']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Mean', 'Other']

gingOpenendedVQABenchmarking2024

Open-ended VQA benchmarking of Vision-Language models by exploiting Classification datasets and their semantic hierarchy

Include

null

This paper propose a novel VQA benchmark based on well-known visual classification datasets which allows a granular evaluation of text-generative vision-language models and their comparison with discriminative vision-language models. To improve the assessment of coarse answers on fine-grained classification tasks, the authors suggest using the semantic hierarchy of the label space to ask automatically generated follow-up questions about the ground-truth category.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Open-ended visual understanding

No

null

Comprehensive

null

Open-ended Visual Question Answering (oVQA), which tests vision-language models (VLMs) on their visual understanding by asking questions via natural language. Unlike multiple choice VQA, where answers can be chosen from a predefined set of options, oVQA requires the model to generate the answer rather than simply choosing the option with the highest score.

Image, question and a golden answer

null

Modified from another benchmark (e.g. translation into another language)

95,864

Yes

attribute type

Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), - Contains: A less restrictive option is to consider a response correct if the prediction contains the true class name after preprocessing - ClipMatch: matching the prediction and label using cosine similarity in a vector embedding space

null

Academia

Yes

null

Test

null

Simple Mean

Yes

Stratification of results based on the used VQA benchmarks

null

https://github.com/lmb-freiburg/ovqa

OVQA

Not defined

There is no specified phenomenon besides the models' ability to answer open-ended questions.

Yes

No

No comparisons made

No

null

simple mean and variance

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Representative task (e.g. answering medical licensing exam questions)

null

Authors' description is unclear

Not applicable

null

No

VQA

null

['Another benchmark']

['Criterion']

['Short free response']

['Exact match', 'Soft match']

['No definition']

['No']

['Yes']

['No comparison made']

['No']

['Partial', 'Representative']

['Mean', 'Std']

hwangMultitaskBenchmarkKorean2022

A Multi-Task Benchmark for Korean Legal Language Understanding and Judgement Prediction

Include

null

This work introduces LBOX OPEN, the first large-scale benchmark of Korean legal AI datasets, comprising a legal corpus of 147k precedents and multiple tasks including classification, legal judgment prediction, and summarization. It also presents LCUBE, the first Korean legal language model trained on this corpus.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

legal language understanding and legal judgment prediction in the Korean context

Yes

we release LBOX OPEN, the first large-scale Korean legal AI benchmark that consists of six datasets: (1) a large-scale legal precedent corpus (PRECEDENT CORPUS), (2) two classification tasks (CASE NAME, STATUTE), (3) two legal judgement prediction tasks (LJP-CRIMINAL, LJPCIVIL), and (4) one summarization task (SUMMARIZATION).

Subset

null

legal text classification (predicting case names and statutes from factual case descriptions), legal judgment prediction (estimating punishment ranges or claim acceptance levels from case facts), and summarization (generating summaries of legal rulings and reasoning sections).

input text and the corresponding label or output in case of summarisation

null

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Expert-crafted task examples (e.g. hand-written examples)

14.1k

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train, Validation

147k precedents (259M tokens)

null

Simple Mean

No

null

https://github.com/lbox-kr/lbox-open

LBOX OPEN

Contested

Yes

No

No comparisons made

No

Yes

They discuss how the tasks are grounded in real-world legal processes, and mention that legal judgment prediction tasks remain especially challenging

simple mean and std

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

it is based on the real document, but not used in the real cases, and tasks as summarisation and classification are the constructed ones

Composite phenomenon

Yes

null

No

Law

null

['Real task', 'Author-crafted', 'Expert-crafted']

['Convenience', 'Targeted', 'Criterion']

['Short free response', 'Free response']

['Exact match', 'Soft match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Partial', 'Constructed']

['Mean', 'Std']

shiLargeLanguageModels2023

Large Language Models Can Be Easily Distracted by Irrelevant Context

Include

null

The paper introduces GSM-IC, a variant of the GSM8K arithmetic reasoning dataset that includes irrelevant sentences to test large language models' distractibility. The authors evaluate several prompting strategies on LLMs, revealing significant performance drops when irrelevant information is present, and explore mitigation strategies including self-consistency decoding, exemplar design, and explicit instructions that partially restore performance.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

reasoning, distractibility (irrelevant context QA)

Yes

Filtering out irrelevant information is essential for handling real-world tasks. Our evaluation indicates that despite the strong performance on challenging reasoning problems, state-of-the-art language models still have fundamental weaknesses in context understanding and identifying the relevant information from the input. Our findings suggest that in order to gain a more holistic understanding of the reasoning capability of language models, future work should also consider the model sensitivity to irrelevant context, in addition to solving more challenging problems. (pg 2.)

Subset

null

Given a maths word problem containing one irrelevant sentence, models are expected to output the numeric answer.

Each item = (Problem and Answer). The problems are derived from GSM8K and distractor sentences.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language), Procedurally-generated task examples (e.g. Creating instances from a template)

4000 problems

Yes

Reasoning steps, distractor categories (topic/ overlap/ number range)

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

First, derive from existing benchmark (GSM8K), Then, create distractors via templates. Finally, manually verify grammaticality and answer invariance.

Mix (multiple authors from industry and academia)

Yes

null

Test, Train, Validation

100 problems for validation.

Answers are expected to be exact integers; authors mark incorrect if formatting differs.

Simple Mean

Yes

Accuracies by reasoning step or by distractor category

null

https://github.com/google-research-datasets/GSM-IC

GSM‑IC

Contested

Yes

No

Yes

Validty factors (topic, overlap, number rage) and the discussion for construct limitations show that the task isolates the distractability (irrelevant context), which authors address as the current limitation of the reasoning.

Percentage, comparison

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

Because the distractor sentences are from controlled templates, the new benchmark may nt be realistic due to the superfluous information built from synthetic perturbations.

Single cohesive phenomenon

Not applicable

58052 problems total but 4000 instances randomly drawn from the full set as an evaluation subset for cheaper computational cost.

null

Reasoning

Mathematical

null

['Author-crafted', 'Another benchmark', 'Procedurally-generated']

['Convenience', 'Targeted']

['Short free response']

['Exact match']

['Contested']

['Yes']

['Comparison made']

['Yes']

['Constructed']

['Mean']

leeVHELMHolisticEvaluation2024

VHELM: A Holistic Evaluation of Vision Language Models

Include

null

The paper's main contributions are three-fold. First, the authors identify the aspects that are both applicable to VLMs and important to evaluate from either a technological or societal perspective. Second, they assemble 21 existing VLM benchmark datasets, which are sets of image-text prompts and expected output, and map to the aspects to ensure complete coverage. Third, they standardize the evaluation procedures so that apple-to-apple comparisons can be made across the models.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

visual understanding

Yes

This benchmark measures multiple visual phenomena in VLMs and more specifically: bias, fairness, knowledge, multilinguality, reasoning, robustness, toxicity, and visual perception

Comprehensive

null

A scenario represents a use case for a VLM and is identified by a task (e.g., question answering, code generation, and captioning) and a usage category such as the domain, origin, language, or subject. An example scenario is “visual question answering on medical images” where the task is visual question answering and the usage category is medical images.

the aspect of the task, a prompt, an image, a gold response and the corresponding metric reference for this task

null

Modified from another benchmark (e.g. translation into another language)

9,000

Yes

aspect of the task and the respective metrics to be used

Convenience sample (creators found a set of tasks that was readily accessible)

Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics), LLM post-processing (extracting answers, reformatting for automated scoring), Win rate

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

The results are stratified based on the used benchmarks

null

https://crfm.stanford.edu/helm/vhelm/v2.0.1/

VHELM

Contested

Yes

The benchmark is itself realistic

No

null

simple mean

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

No

VQA

null

['Another benchmark']

['Convenience']

['Short free response', 'Free response']

['Exact match', 'Soft match', 'LLM-as-a-Judge', 'LLM post-processing', 'Reward']

['Contested']

['Yes']

['Realistic']

['No']

['Representative']

['Mean']

wangM4GTbenchEvaluationBenchmark2024

M4GT-Bench: Evaluation Benchmark for Black-Box Machine-Generated Text Detection

Include

null

M4GT-Bench, a multilingual, multi-domain benchmark for detecting machine-generated text (MGT). The benchmark extends the previous dataset and contains three distinct tasks: Binary MGT Detection (classifying text as human-written or machine-generated), multi-way Generator Detection (which LLM generated text), change Point Detection (text transitions from human-written to machine-generated).

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Machine-generated text (MGT) detection

Yes

Machine-generated text detection as the task of identifying and differentiating LLM-generated text from genuine human-generated text.

Subset

null

Benchmark tasks: (1) Binary classification to determine if text is human-written or machine-generated, (2) Multi-way classification to identify which specific LLM generated a text, and (3) Boundary detection to identify where text transitions from human-written to machine-generated content.

Text sample, label human-written/machine-generated, model (if applicable)

null

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

73,288 (total machine)

Yes

Domain; language; generating LM; boundary position

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice, Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), Mean Absolute Error

null

Academia

Yes

null

Test, Train, Validation

Mixed Dataset, PeerRead, e.g. ChatGPT as Generator: 3,649 train; 1,522 test; 505 dev

null

Simple Mean

Yes

LM generator, language, domain

null

https://github.com/mbzuai-nlp/M4GT-Bench

M4GT-Bench

Widely-agreed

Yes

No

Yes

Human evaluation for multi-way generator detection, where they test how well humans can distinguish between different LLM generators.

Simple means for the main metrics

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

Represent realistic scenarios where one would want to detect fully machine-generated text, identify which model generated text, or find the point where human writing ends and machine generation begins.

Composite phenomenon

Yes

null

No

NLP

null

['Real task', 'Another benchmark', 'LLM-generated']

['Convenience', 'Targeted']

['Multiple choice', 'Short free response']

['Exact match', 'Soft match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Mean']

comsaBenchmarkReasoningSpatial2023

A Benchmark for Reasoning with Spatial Prepositions

Include

null

The paper introduces a new benchmark designed to evaluate the inferential capabilities of statements involving spatial prepositions. Featuring original datasets in both English and Romanian, the benchmark explores the boundaries of large language models’ reasoning about spatial relations.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Spatial reasoning

Yes

The benchmark proposes the challenge of using spatial prepositions to refer to abstract concepts in addition to physical relationships.

Subset

null

Take two premises with prepositions and determine if the conclusion holds.

Two premies, one conclusion, whether the conclusion holds.

null

Expert-crafted task examples (e.g. hand-written examples)

800

Yes

The preposition used in a premise / conclusion in an example.

Targeted items (creators defined a task space and chose tasks within it strategically)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Industry

Yes

null

Test

null

Yes

items containing different prepositions

null

https://github.com/google-research/language/tree/master/language/spatial_prep

null

Contested

Yes

No

No comparisons made

No

Yes

No

null

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

No

Reasoning

Spatial

null

['Expert-crafted']

['Targeted']

['Short free response']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['No']

['Constructed']

null

luoCODISBenchmarkingContextdependent2024

CODIS: Benchmarking Context-dependent Visual Comprehension for Multimodal Large Language Models

Include

null

This paper introduces the CODIS benchmark for evaluating MLLMs on their ability to incorporate free-form textual context to improve image understanding. The authors show that current MLLMs underperform compared to humans and struggle to effectively extract and utilize contextual information to improve their understanding of images.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

context-dependent visual comprehension

Yes

understand the extent to which MLLMs can leverage context to enhance their visual comprehension

Subset

null

Take an image and question, and provide a short-form answer

<image> Question: When was this photo probably taken, the first or second half of the year? Context: I took this photo when I was in Australia. Ground-truth Answer: First half

null

Expert-crafted task examples (e.g. hand-written examples)

706

Yes

Taxonomy of Context

Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

They report accuracy as the main evaluation metric, assessed by humans for the primary results. They also evaluate LLM-as-judge for accuracy measurement in ablation study

null

Academia

Yes

null

Test

null

Simple Mean

Yes

null

https://github.com/THUNLP-MT/CODIS

CODIS

Contested

Yes

No

No comparisons made

No

Yes

Five annotators participated in the data collection process. To ensure the quality of our dataset, each submission by an annotator was cross-checked by the other four.

simple mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Single cohesive phenomenon

Not applicable

null

No

VQA

null

['Expert-crafted']

['Criterion']

['Short free response']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Partial']

['Mean']

lalCaTbenchBenchmarkingLanguage2024

null

Include

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

(causal/ temporal) reasoning

Yes

While LLMs appear to generate good plans, it’s unclear how well they understand important aspects of the steps themselves. We thus use CAT-BENCH to test whether LLMs can identify step dependencies that reflect the causal and temporal structure of the plan. We find that current LLMs struggle to identify step dependencies, often performing close to random chance, raising more questions about their understanding of instructional text. (Section1)

Subset

Faithfulness in explanation qualities

A step order prediction with plan text and question of how many step happens before/after the certain step. The output is expected to be yes/no binary answer with optional explanation generation.

One binary question about a specific order pair of steps and ground-truth label

null

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

4260

Yes

Recipe, temporal step, causal dependency

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice, Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), Human ratings (text quality, preference, NOT manual scoring of other metrics)

null

Academia

Yes

null

Test

null

Weighted Mean

Yes

Analysis by class, relation direction (before/after), step distance (close/distant)

null

https://github.com/StonyBrookNLP/CaT-Bench

CaT-Bench: Benchmarking Language Model Understanding of Causal and Temporal Dependencies in Plans

Contested

Yes

No

Yes

High inter‑annotator agreement and consistency metrics

Means, standard deviations, weighted Fleiss‑k

Outputs alone

Representative task (e.g. answering medical licensing exam questions), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

End‑users would not normally answer yes/no (binary) dependency questions so it is constructed task. Plus, the benchmark captures a key capability (causal/temporal reasoning over plans) that underlies practical applications such as recipe‑guidance, robotics and agent planning, so it is representative of real‑world needs even though the evaluation setting is synthetic.

Composite phenomenon

No

2840 (balanced subset for evaluation) + 1420 (held‑out non‑dependent questions used for analysis)

Yes

Reasoning

Planning

null

['Real task', 'Author-crafted']

['Convenience', 'Targeted']

['Multiple choice', 'Free response']

['Exact match', 'Human ratings']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Representative', 'Constructed']

['Mean', 'Std']

suActPlan1KBenchmarkingProcedural2024

ActPlan-1K: Benchmarking the Procedural Planning Ability of Visual Language Models in Household Activities

Include

null

This paper introduces ActPlan-1K for evaluating VLMs on procedural and counterfactual reasoning tasks. By combining natural language descriptions with simulated environment images, the benchmark assesses the ability of VLMs to generate coherent action plans.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

multimodal reasoning, procedural planning

Yes

Whether VLMs can generate plausible action plans for multi-modal embodied AI task

Subset

null

Given household environment E, there are manipulable objects set O. For each household activity T , an VLM agent A takes in the task description T and environment images {I1, I2, ...} as input, generates procedural plan P∗ that can accomplish the task. The household environments have multiple interior spaces therefore there are multiple images to ensure that necessary spatial information is provided.

NL Task description: "The task goal is to assemble gift baskets. .... <description of the environment>" Gold Plan

null

Expert-crafted task examples (e.g. hand-written examples)

1000

Yes

normal or counterfactual

Targeted items (creators defined a task space and chose tasks within it strategically)

Structured response (e.g. valid JSON, API call alone)

n-gram (BLEU, ROUGE, chrF), Human ratings (text quality, preference, NOT manual scoring of other metrics)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

null

https://github.com/HKUST-KnowComp/ActPlan-1K

ActPlan-1K

Widely-agreed

Yes

No

No comparisons made

No

null

Outputs alone

Representative task (e.g. answering medical licensing exam questions), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

No

Reasoning

Planning

null

['Expert-crafted']

['Targeted']

['Structured']

['Soft match', 'Human ratings']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Representative', 'Constructed']

null

yinGeoMLAMAGeodiverseCommonsense2022

GeoMLAMA: Geo-Diverse Commonsense Probing on Multilingual Pre-Trained Language Models

Include

null

In this paper, a benchmark dataset is introduced, Geo-diverse Commonsense Multilingual Language Models Analysis (GEOMLAMA), for probing the diversity of the relational knowledge in multilingual PLMs.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

geo-diverse commonsense probing

Yes

do PLMs store geo-diverse commonsense knowledge?

Subset

null

investigating whether PLMs are capable of predicting correct answers among all the possibilities of different countries.

A prompt and a gold answer. For each concept, there are multiple masked multilingual prompts with specified country information [X] querying geo-diverse knowledge about the concept.

null

Expert-crafted task examples (e.g. hand-written examples)

3125

Yes

Language of the prompt.

Random sample (creators defined a task space and sampled from it)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), Distribution (perplexity, calibration, correlation)

null

Academia

Yes

null

Yes

By language.

null

https://github.com/WadeYin9712/GeoMLAMA

GEOMLAMA

Contested

Yes

No

No comparisons made

No

null

Model access required (e.g. logits)

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

No

Knowledge

Cultural

null

['Expert-crafted']

['Random']

['Short free response']

['Exact match', 'Distribution']

['Contested']

['Yes']

['No comparison made']

['No']

['Constructed']

null

kesenViLMAZeroshotBenchmark2024

ViLMA: A Zero-Shot Benchmark for Linguistic and Temporal Grounding in Video-Language Models

Include

null

This paper introduces VILMA, a zero-shot benchmark for evaluating VidLMs, designed to require strong temporal understanding. The authros adopt the following methodology: (i) they harvest high-quality examples from existing video-language datasets; (ii) they create counterfactual examples or ‘foils’, so that a test requires distinguishing correct from counterfactual video+text pairs; (iii) they create a proficiency test to gauge if a model learns the capabilities we deem necessary to solve the main test; (iv) they apply automatic and manual validation of the examples and their counterfactuals to control for biases and to ensure a high-quality evaluation benchmark.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

temporal grounding in video-language models

Yes

In principle, VidLMs can visually ground linguistic phenomena which are beyond the reach of image-language models (ILMs), since videos include dynamically evolving phenomena (e.g., events, actions, physical processes). This temporal dimension makes learning more complex.

Comprehensive

null

- The Action Counting task probes the ability of models to accurately count the occurrences of actions within a given video input stream. - The Situation Awareness task shows how effectively VidLMs grasp the interaction between visual clues and verbal context by testing whether they recognise actors, actions, and their relationships. - The Change of State task examines the ability of VidLMs (i) to recognise and distinguish different sub-phases of actions, especially those that induce a change of state (CoS) of objects or entities involved in it; and (ii) to align the beginning and ending phases of these actions across modalities. - The Rare Actions task probes how well VidLMs identify novel compositions and recognise unusual interactions between human beings and objects. - The Spatial Relations task focuses on the ability of models to distinguish different spatial and spatio-temporal relations related to the actions carried out in a video (e.g. moving an object ‘over’, or ‘towards’ another object).

Video frames, ground truth video caption, foil video caption

null

Modified from another benchmark (e.g. translation into another language)

5,934

No

null

Convenience sample (creators found a set of tasks that was readily accessible)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

results are stratified based on the temporal dimension/temporal task

null

https://github.com/ilkerkesen/ViLMA/tree/main

ViLMA

Widely-agreed

Yes

No

No comparisons made

The benchmark is itself realistic

No

Yes

They manually checked every video-caption-foil sample, retaining only those in which the foil was unambiguously false with respect to the input video. This resulted in the removal of 1278 (15.11%) of samples in the proficiency tests. The main tests were validated independently, in a study conducted on AMTurk.

simple mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

No

Grounding

null

['Another benchmark']

['Convenience']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['Realistic']

['Yes']

['Partial']

['Mean']

kurticMathadorLMDynamicBenchmark2024

Mathador-LM: A Dynamic Benchmark for Mathematical Reasoning on Large Language Models

Include

null

The paper introduces Mathador-LM, a arithmetic reasoning benchmark where models must reach a target number using five input numbers and basic arithmetic operations. Evaluations across various LLMs show that even top models perform far below the level of 3rd-grade students, revealing significant gaps in mathematical reasoning while avoiding test-data contamination issues.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

reasoning, planning, math

Yes

LLM performance of interpreting a ruleset and planning a valid sequence of arithmetic is critical for their reasoning evaluation. We propose an alternative pathway towards reliable examination of LLM performance via dynamic, one-time benchmarks that mitigate contamination by being created on-the-fly, independently for each evaluation run.. This approach mitigates issues such as test-set leakage into training data and provides a reliable method to evaluate closed-source models, even in the absence of detailed information about their training data.

Subset

null

Given 5 base numbers and a target, generate up to 4 arithmetic steps that yield the target while obeying game constraints.

Target number, a list of 5 base numbers, a 4‑line expression sequence as a response.

null

Procedurally-generated task examples (e.g. Creating instances from a template)

1000 items

Yes

Difficulty (calculated by the average attainable score across all solutions for the target)

Random sample (creators defined a task space and sampled from it), Specific criteria (items were taken from a larger set based on specified rules)

Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall)

Normalised game‑score accuracy (to yield 0-100%)

All items come from an algorithmic generator that samples numbers within predefined ranges and verifies solvability. There are no hand‑written examples, no crowd‑sourcing, no exam questions, and no LLM‑generated prompts.

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

Difficulty level, error type

null

https://github.com/IST-DASLab/Mathador-LM

Mathador-LM

Widely-agreed

Yes

No

Yes

stability study across different regenerated datsets showing low variance

simple mean, 95% confidence interval, percentage point of performance gains over baselines

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

No

null

Yes

Reasoning

Mathematical

null

['Procedurally-generated']

['Random', 'Criterion']

['Free response']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean', 'Tests']

zhangCABComprehensiveAttention2023

CAB: Comprehensive Attention Benchmarking on Long Sequence Modeling

Include

null

CAB is a multimodal benchmark assessing long-range modeling in transformers across computer vision, natural language processing, speech processing, and time-series forecasting. It is publicly available and composed of 7 tasks, spanning 9 datasets, to measure noncausal self, causal self, noncausal cross, and causal cross attention with a custom metric.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Long-range modeling

Yes

Long range modeling is implicitly defined as "longer sequence modeling in different domains" (1).

Subset

The four types of attention to be measured are noncausal self attention, causal self attention, noncausal cross attention, causal cross attention.

There are "seven tasks covering four research fields ... computer vision, natural language processing, speech processing, and time series forecasting" (3). The tasks are Text-to-Speech Synthesis (TTS), Summarization (Sum), Long Sequence Time-series Forecasting (LSTF), Point Cloud Completion (PCC), Langauge Modeling (LM), Masked Language Modeling (MLM), Super-Resolution (SR).

The benchmark is composed of 9 distinct datasets with different features.

Each task has its own dataset and metric.

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language)

null

No

null

Unknown

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF), LLM post-processing (extracting answers, reformatting for automated scoring), Distribution (perplexity, calibration, correlation), Mel-Cepstral Distortion is a measure of audio quality for TTS. A custom index is defined to balance all the evaluation metrics.

TTS uses MCD, MSD. Sum uses ROUGE. LSTF uses MSE, MAE. PCC uses CD and F-Score. LM and MLM uses PPL. SR uses PSNR and SSMI. The paper defines a custom metric, compositional index (CI), which is "a normalized score to balance the influence among evaluation metrics, and high CI represents excellence. It is computed as follows: a) we transform all evaluation metrics beforehand, so that a higher score indicates better performance; b) we then normalize each transformed metric with Z-score normalization; c) after normalization, the score of each evaluation metric is averaged within each task, and is further averaged across tasks" (5).

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Custom compositional index (CI)

Yes

Scores are reported for each attention type, per task, per sub-metric in each task, and with the total CI.

null

https://github.com/Shark-NLP/CAB

CAB (Comprehensive Attention Benchmark)

Widely-agreed

Yes

No

No comparisons made

The benchmark is itself realistic

No

Yes

The authors highlight that evaluating long-range modeling requires assessing "standard bidirectional (or noncausal) self attention" and "cross attentions and unidirectional (or causal) attentions, which are equally important to downstream applications" (1). CAB is proposed to measure causal and cross self attention, in addition to noncausal self attention.

Simple mean/sum, custom normalized aggregate metric

Model access required (e.g. logits)

Complete real task (e.g. providing medical advice to real people interactively)

null

Composite phenomenon

Yes

Each task has a different target sequence length.

No

NLP

Long Context

null

['Real task', 'Another benchmark']

['Unknown']

['Multiple choice', 'Short free response', 'Free response']

['Exact match', 'Soft match', 'LLM post-processing', 'Distribution', '']

['Widely-agreed']

['Yes']

['Realistic']

['Yes']

['Complete']

['Mean', 'Other']

chenMLLMasajudgeAssessingMultimodal2024

MLLM-as-a-Judge: Assessing Multimodal LLM-as-a-Judge with Vision-Language Benchmark

Include

null

This paper introduces the MLLM-as-a-Judge benchmark to assess the ability of MLLMs in assisting judges across diverse modalities, encompassing three distinct tasks: Scoring Evaluation, Pair Comparison, and Batch Ranking.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

MLLM-as-a-judge

null

capability of MLLMs in tasks of Scoring Evaluation, Pair Comparison and Batch Ranking.

Comprehensive

null

Take a single MLLM response, provide the score; Take two MLLM responses, compare which one is better; Take a batch of MLLM responses, provide a ranking

<two MLLM responses>, Judgement: B

null

Expert-crafted task examples (e.g. hand-written examples)

17903

Yes

input setting

Targeted items (creators defined a task space and chose tasks within it strategically)

Short free response (e.g. single word or number)

Correlation (Matthew's correlation, Pearson's r)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

null

https://github.com/Dongping-Chen/MLLM-Judge

MLLM-as-a-Judge

Widely-agreed

Yes

null

Yes

No

No comparisons made

No

Yes

we implement cross-validation between different annotators and conduct continuous monitoring to ensure they are maintaining objectivity and fairness.

null

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

No

LLM as a Judge

null

['Expert-crafted']

['Targeted']

['Short free response']

['Correlation']

['Widely-agreed']

['Yes']

['']

['No comparison made']

['Yes']

['Partial', 'Constructed']

null

gandhiUnderstandingSocialReasoning2023

Understanding Social Reasoning in Language Models with Language Models

Include

null

The paper introduces BigToM, a comprehensive benchmark containing 5,000 Theory-of-Mind scenarios created through procedural generation using GPT-4-populated causal templates and validated by human raters. This benchmark addresses key limitations in existing ToM evaluations of large language models, specifically the inconsistent results across previous studies and methodological concerns about evaluation validity. This benchmark aims to provide a more rigorous framework for assessing how well AI systems can understand human mental states.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Theory-of-mind, (social) reasoning

Yes

Theory of Mind (ToM) is defined as the ability to attribute latent mental states (beliefs, desires, knowledge, emotions) to agents and use them to explain or predict behavior. By representing ToM scenarios as causal graphs, we can systematically intervene on variables, generate control conditions, and probe different aspects of an LLM’s ToM capabilities.

Subset

null

Multiple‑choice comprehension questions where a model is expected to infer an agent’s belief or action given a short story generated from a causal template.

Story, Question, 2 answer options, Correct answer, Condition category.

null

Procedurally-generated task examples (e.g. Creating instances from a template), LLM-generated task examples (e.g. Filtered from responses to a prompt)

5000 items

Yes

Condition category (e.g., True/False Belief in terms of theory-of-mind)

Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

Condition category

null

https://sites.google.com/view/social-reasoning-lms

BigToM

Contested

Yes

No

Yes

expert review + participant rating studies, coherence test

Mean with 95% Confidence Interval

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

Synthetic narratives rather than real‑world usage, designed for controlled probing.

Composite phenomenon

Yes

null

Yes

Theory of Mind

null

['Procedurally-generated', 'LLM-generated']

['Criterion']

['Multiple choice']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean', 'Tests']

bandarkarBelebeleBenchmarkParallel2024

The BELEBELE Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants

Include

null

BELEBELE is a multiple-choice machine reading comprehension benchmark designed to evaluate language models' multilingual capabilities. It's covering diverse languages and scripts from high-, medium-, to low-resource languages.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

reading comprehension, multilingual

Yes

Multilingual reading comprehension as the ability to understand text passages and correctly answer multiple-choice questions about those passages across different languages.

Subset

null

Multiple-choice reading comprehension: LMs read a passage and answer a question about it by selecting the correct option from four possible answers.

passage from FLORES-200, question about the passage, four multiple-choice answers

null

Crowd-sourced task examples (e.g. Prolific-created tasks), Modified from another benchmark (e.g. translation into another language)

67.5k training samples, 3.7k development samples

Yes

language, script, language family

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Adapted FLORES-200 machine translation dataset. Questions and answers were created in English by professional annotators.

Academia

Yes

null

Test, Train

null

Simple Mean

Yes

Per-language scores, Scores by language family, Scores by script type

null

https://github.com/facebookresearch/belebele

BELEBELE

Widely-agreed

Yes

No

Yes

Statistical t-tests to ensure the distribution of features; Training a logistic regression; Comparison with human performance; Cross-correlation with established benchmarks

Simple mean accuracy

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Single cohesive phenomenon

Not applicable

null

Yes

NLP

Understanding

Multilinguality

['Crowd-sourced', 'Another benchmark']

['Targeted', 'Criterion']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Mean']

panchalWhatSayWhen2024

What to Say and When to Say it: Live Fitness Coaching as a Testbed for Situated Interaction

Include

null

Open-ended, asynchronous interactions, where an AI model may proactively deliver timely responses or feedback based on the unfolding situation in real-time, are an open challenge. This work presents the QEVD-FIT-COACH benchmark and dataset, which explores human-AI interaction in the challenging, yet controlled, real-world domain of fitness coaching – a task which intrinsically requires monitoring live user activity and providing immediate feedback.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Situated interaction

Yes

A notable type of situated interaction is the instructional or coaching scenario, where an instructor guides a user through a complex activity, such as live fitness coaching.

Subset

null

Feedback Structure: The feedbacks in the QEVD-FIT-COACH benchmark have the following structure: At the start of each exercise, acknowledging feedback is given once the user has started; otherwise, a reminder to do so is provided. A corrective feedback is provided as soon as a mistake is clearly visible. Similarly, when the user begins to correct their mistake, feedback is provided to acknowledge and guide the user to successfully correct the error. If the user is performing the exercise correctly, feedback focuses on repetition counting. Finally, at the end of each exercise, a feedback focused on the overall performance during that exercise is provided.

Video frames, a list of feedback statements that correspond to a specific timestep

null

Real task examples (e.g. GitHub issues), LLM-generated task examples (e.g. Filtered from responses to a prompt)

28,326

No

null

Random sample (creators defined a task space and sampled from it)

Free response (e.g. summary paragraph, executable code), Extended interaction (e.g. conversation, calling an API and processing the response)

n-gram (BLEU, ROUGE, chrF), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train

377,678

null

Simple Mean

No

null

https://github.com/Qualcomm-AI-research/FitCoach

QEVD-FIT-COACH

Contested

Yes

No

No comparisons made

The benchmark is itself realistic

Yes

No

null

simple mean

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

null

Single cohesive phenomenon

No

null

No

User Interaction

null

['Real task', 'LLM-generated']

['Random']

['Free response', 'Interaction']

['Soft match', 'LLM-as-a-Judge']

['Contested']

['Yes']

['Realistic']

['No']

['Complete']

['Mean']

zhangCABComprehensiveAttention2023

CAB: Comprehensive Attention Benchmarking on Long Sequence Modeling

Include

null

Current benchmarks testing different attention architectures for long-term modelling only focuses on the standard bidirectional (or noncausal) self attention, and completely ignores cross attentions and unidirectional (or causal) attentions. In this paper, we propose Comprehensive Attention Benchmark (CAB) with four distinguishable attention patterns, namely, noncausal self, causal self, noncausal cross, and causal cross attentions. In seven tasks, CAB validates efficient attentions in eight backbone networks to show their generalization across neural architectures.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Long sequence modelling capability of transformers with different attention mechanisms

No

It is defined as the performance of the transformer models in handling long-sequence tasks. There are seven tasks, such as text-to-speech synthesis, summarization, Long Sequence Time-series Forecasting, etc. The sequence length considered 'long' would be different for each task.

Comprehensive

null

There are seven long-sequence tasks ranging from Computer Vision, NLP and time-series forecasting handled by transformers. Examples include Super-Resolution, Masked Language Modelling, Long Sequence Time-series Forecasting.

A long-sequence task, the dataset for the task, sequence length, evaluation metric for performance, transformer model type, attention mechanism

null

Real task examples (e.g. GitHub issues)

null

Yes

The backbone transformer architecture and attention architecture

Convenience sample (creators found a set of tasks that was readily accessible)

Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF), Distribution (perplexity, calibration, correlation), MCD, MSD, PSNR, SSIM

Different metrics used to evaluate each task.

null

Academia

Yes

null

Test

null

Different outputs for each task. For example, for summarisation, it produces a summary paragraph; for Super-Resolution, it convert low-resolution (16 × 16) face images into high-resolution (128 × 128) images.

null

Yes

Results are provided separately for each task with different metrics.

null

https://github.com/Shark-NLP/CAB

Comprehensive Attention Benchmark (CAB)

Not defined

Yes

No

No comparisons made

The benchmark is itself realistic

No

null

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

null

Composite phenomenon

Yes

So one dataset for each task, seven tasks in total

No

NLP

Long Context

null

['Real task']

['Convenience']

['Free response']

['Exact match', 'Soft match', 'Distribution', 'Soft match']

['No definition']

['Yes']

['Realistic']

['No']

['Complete']

null

tianDiagnosingFirstorderLogical2021

Diagnosing the First‑Order Logical Reasoning Ability Through LogicNLI

Include

null

The paper introduces LogicNLI, a NLI‑style benchmark crafted to diagnose large language models’ first‑order logic (FOL) reasoning abilities. It disentangles logic from commonsense by procedurally generating facts, rules and statements covering seven FOL operators, and evaluates models along four axes: accuracy, robustness, generalization and proof‑based traceability. Experiment results show substantial gaps to human performance and highlight weaknesses in negation and universal quantification.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

(logical) reasoning

Yes

Authors define FOL reasoning as multi‑step inference using the seven foundational logical operators applied to simple propositions expressed in natural language.

Subset

null

Given a set of natural‑language facts and rules, model is expected to predict whether a hypothesis is entailment, contradiction, neutral, or paradox relative to the premise.

premise texts (facts + rules), hypothesis statement, groud-truth label, proofs

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Procedurally-generated task examples (e.g. Creating instances from a template)

2000

Yes

label balance, hop counts, subject/predicate/vocab sizes, sequence lengths

Random sample (creators defined a task space and sampled from it), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

No, link is broken

null

Test, Train, Validation

16000 (train), 2000 (validation)

null

Simple Mean

Yes

robustness, generalization, traceability

null

https://github.com/tianyikillua/LogicNLI

LogicNLI

Contested

Yes

No

Yes

complementary test suits and analyses per operator

simple mean

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

No

Reasoning

Logical

null

['Author-crafted', 'Procedurally-generated']

['Random', 'Criterion']

['Multiple choice']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean']

zhouRICAEvaluatingRobust2021

RICA: Evaluating Robust Inference Capabilities Based on Commonsense Axioms

Include

null

RICA is a benchmark for evaluating language models' ability to make robust commonsense inferences despite linguistic variations and logical perturbations. Starting from first‑order‑logic templates that encode commonsense axioms, the authors automatically crawl ConceptNet and ATOMIC, then apply 24 perturbation types (negation, antonym, paraphrase, etc.). Experiments show that pre-trained language models perform poorly on robust inference tasks even after fine-tuning, highlighting a significant gap between current AI systems and human-level understanding.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

natural language inference (for commonsense knowledge), first-order logic

Yes

Authors formalize each commonsense axiom as a first‑order‑logic implication and deem a model successful only if it solves all logically‑equivalent probes for that axiom

Subset

null

To test language models' commonsense reasoning and robustness to linguistic variations, the authors transform first-order logic axioms into multiple syntactically different statements expressing the same inferential relationship, then evaluate models through task-specific probes, considering models successful only if they perform like humans across all variations of the axioms.

Premise + conclusion statement (+ with a [MAST] token) + a pair of novel entity strings

null

Crowd-sourced task examples (e.g. Prolific-created tasks), Procedurally-generated task examples (e.g. Creating instances from a template)

2600

No

null

Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice, Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Academia

No, link is broken

null

Test, Train, Validation

8000 (train)

null

Simple Mean

No

null

https://sites.google.com/usc.edu/rica

RICA

Contested

Yes

No

Yes

Ablation studies, perturbation analysis, knowledge control

Simple mean + standard deviations

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

2600 testsets (joint human‑verified + curated), 8000 training sets (plus optional 100,000 noisy or 257,000 raw datasets)

null

Reasoning

Logical

null

['Crowd-sourced', 'Procedurally-generated']

['Targeted']

['Multiple choice', 'Short free response']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean', 'Std']

agrawalLargeLanguageModels2022

Large Language Models are Few-Shot Clinical Information Extractors

Include

null

Datasets for benchmarking LLMs on five different clinical NLP tasks: clinical sense disambiguation, biomedical evidence extraction, coreference resolution, medication status extraction, and medication attribute extraction. They show that LLMs perform well on these tasks despite not being specifically trained for the clinical domain.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

clinical information extraction

Yes

Clinical information extraction as "the extraction of important variables trapped in clinical notes" and structuring clinical variables from unstructured text, (handling ambiguous jargon and nonstandard phrasal structure specific to clinical text).

Subset

null

Tasks in clinical information extraction: 1) clinical sense disambiguation, 2) biomedical evidence extraction, 3) coreference resolution, 4) medication status extraction, and 5) medication attribute extraction.

For example, in clinical sense disambiguation, an item consists of a clinical note and an abbreviation to be expanded.

null

Real task examples (e.g. GitHub issues), Expert-crafted task examples (e.g. hand-written examples)

null

Yes

medication type, status, relation types, antecedent-pronoun pairs

Convenience sample (creators found a set of tasks that was readily accessible), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Clinical text from primarily the CASI dataset (Clinical Acronym Sense Inventory) and creates new annotations on this text.

Academia

Yes

null

Test, Train, Validation

null

Simple Mean

Yes

Scores for each of the five tasks.

null

https://huggingface.co/datasets/mitclinicalml/clinical-ie

null

Widely-agreed

Yes

No

null

Primary metrics used are F1 scores, accuracy, recall, and precision. Partially, micro and macro averages are reported.

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

Yes

NLP

Extraction

Medicine

['Real task', 'Expert-crafted']

['Convenience', 'Criterion']

['Short free response', 'Structured']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Partial', 'Constructed']

['Mean']

guhaLegalBenchCollaborativelyBuilt2023

LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models

Include

null

LEGALBENCH is a comprehensive benchmark for evaluating language models' legal reasoning capabilities, consisting of 162 tasks spanning six distinct categories of legal reasoning that were designed and hand-crafted by legal professionals. The benchmark provides evaluation code, prompts, and a common vocabulary bridging legal frameworks and AI development, enabling rigorous assessment of both open-source and commercial language models on practically useful legal reasoning skills.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

(legal) reasoning

Yes

Six reasoning types: issue‑spotting, rule‑recall, rule‑application, rule‑conclusion, interpretation, rhetorical‑understanding.

Subset

null

input text (from single sentences to two‑page documents) + prompt + label (classification/ generation).

one legal question, text snippet, gold answer/label.

null

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language), Procedurally-generated task examples (e.g. Creating instances from a template), LLM-generated task examples (e.g. Filtered from responses to a prompt)

91000

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

LEGALBENCH was constructed from a mix of existing legal datasets (restructured for the few-shot LLM paradigm), and hand-crafted datasets created and contributed by legal professionals (included as authors on this work).

Mix (multiple authors from industry and academia)

Yes

null

Test

null

LEGALBENCH tasks also span different formats: multiple-choice questions (35 tasks), open-generation (7 tasks), binary classification (112 tasks), and multi-class/multi-label classification (8 tasks). Tasks range from 50–2 000 samples (avg. approximately 563 samples per task)

Weighted Mean

Yes

162 individual task scores (task structure/reasoning types/legal domains/language variation)

null

https://huggingface.co/datasets/nguha/legalbench

LEGALBENCH

Contested

Yes

No

The benchmark is itself realistic

No

Partially; addressed their own limitations.

We note that the purpose of this work isn’t to evaluate whether computational systems should replace lawyers and legal officers, or to understand the positive and negative impacts of that replacement. Our goal is to construct artifacts that enable the relevant stakeholders and affected communities to better understand, empirically, the capacity for LLMs to perform different types of legal tasks. Given the proliferation of computational legal tools, we believe that answering this question is vital for ensuring their safe and ethical usage. (section 1)

Mean and Standard deviation

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

Clause classification, supply-chain disclosure compliance, privacy-policy entailment

Composite phenomenon

Yes

null

Law

null

['Real task', 'Another benchmark', 'Procedurally-generated', 'LLM-generated']

['Convenience', 'Targeted', 'Criterion']

['Multiple choice', 'Short free response', 'Free response', 'Structured']

['Exact match']

['Contested']

['Yes']

['Realistic']

['Yes']

['Representative']

['Mean', 'Std']

chenToMBenchBenchmarkingTheory2024

ToMBench: Benchmarking Theory of Mind in Large Language Models

Include

null

ToMBENCH is a bilingual benchmark that evaluates language models' Theory of Mind (ToM) capabilities through classic psychology tasks measuring distinct social cognition abilities in a multiple-choice format. Built entirely from scratch to avoid data contamination, the benchmark reveals that even advanced models like GPT-4 still significantly underperform humans in understanding and attributing mental states, particularly when subjected to coherence stress tests. Their aim with ToMBENCH is to enable an efficient and effective evaluation of LLMs’ ToM capabilities, thereby facilitating the development of LLMs with inherent social intelligence.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

reasoning, Theory-of-mind

Yes

null

Comprehensive

Bilingual to avoid training-data leakage and enable the unbiased evaluation.

Given a short social story, a question, and 4 answer options, the model is expected to choose a single correct option.

story, question, option A to D, correct answer

null

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language)

2,860 questions (934 stories) per English/Chinese

Yes

language, task type, ability dimension, story length, annotator agreement

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

scenarios from social media posts

Academia

Yes

null

Test

null

Simple Mean, Weighted Mean

Yes

8 classic psychology tasks, 31 distinct social cognition abilities

null

ToMBench

Contested

Yes

No

Yes

contamination checks, coherence stress test

simple mean with percentage point

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

No

Theory of Mind

null

['Real task', 'Author-crafted', 'Another benchmark']

['Targeted', 'Criterion']

['Multiple choice']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Mean']

wuDetectRLBenchmarkingLLMgenerated2024

DetectRL: Benchmarking LLM-Generated Text Detection in Real-World Scenarios

Include

null

Evaluate LLM-generated text in realistic scenarios. The benchmark collects human-written texts from high-risk domains, generates comparable texts using popular LLMs, and applies various attack methods to simulate real-world conditions.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

LLM-generated text detection

Yes

Defines LLM-generated text detection as the task of discriminating between human-written and LLM-generated texts with a focus on real-world scenarios.

Subset

null

Detect whether text is human-written or LLM-generated, with a focus on real-world scenarios that include various prompt usages, human revisions, and writing noises.

Text sample that is either human-written or generated by an LLM, along with its corresponding label.

null

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Procedurally-generated task examples (e.g. Creating instances from a template), LLM-generated task examples (e.g. Filtered from responses to a prompt)

100,800 human-written samples

Yes

domain, LLM type, attack type, text length

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall), Distribution (perplexity, calibration, correlation)

null

Academia

Yes

null

Test, Train

Task-dependent train and test split (Table 2, p. 4)

Requires a binary classification (human-written vs. LLM-generated) for each text sample.

Simple Mean

Yes

domain, LLM type, attack type, text length intervals

null

https://github.com/NLP2CT/DetectRL

DetectRL

Widely-agreed

Yes

The benchmark is itself realistic

No

Yes

Specifically discuss the validity of their benchmark by comparing it to existing benchmarks and explaining why DetectRL better represents real-world scenarios.

Simple mean to aggregate across different settings. For each detector, they report AUROC and F1 Score values for each specific condition and the average across those conditions.

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

DetectRL simulates realistic conditions under which LLM-generated text detection would need to operate - while not being a complete real task.

Composite phenomenon

Yes

null

No

NLP

Detection

null

['Real task', 'Author-crafted', 'Procedurally-generated', 'LLM-generated']

['Targeted', 'Criterion']

['Multiple choice']

['Exact match', 'Distribution']

['Widely-agreed']

['Yes']

['Realistic']

['Yes']

['Partial']

['Mean']

patelMultiLogiEvalEvaluatingMultistep2024

Multi-LogiEval: Towards Evaluating Multi-Step Logical Reasoning Ability of Large Language Models

Include

null

Multi-LogiEval evaluates language models' multi-step logical reasoning across propositional, first-order, and non-monotonic logic with varied inference rules and depths. Tests on leading models reveal substantial performance drops as reasoning complexity increases, exposing critical gaps in logical reasoning capabilities.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

reasoning, multi-step logic

Yes

The ability to perform multi-step reasoning – drawing conclusions from provided multiple premises -- is a hallmark of human intelligence. Our work aims to bridge these gaps by creating a more comprehensive and logically complex evaluation dataset by incorporating varying numbers of reasoning depths (i.e., multi-steps) to reach conclusions. Our work systematically evaluates multi-hop logical reasoning over various inference rules and their combinations. (section 1)

Subset

null

Given a context (story with logical statements) and a question (candidate conclusion), the model is expected to predict Yes/No. (Binary entailment classification)

Context, question, answer, metadata (e.g., logic type, logic depth)

null

Procedurally-generated task examples (e.g. Creating instances from a template), LLM-generated task examples (e.g. Filtered from responses to a prompt)

1552

No

null

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

15 types (3 logics × 5 depths)

null

https://github.com/Mihir3009/Multi-LogiEval

Multi-LogiEval

Widely-agreed

Yes

null

Yes

No

Yes

Error analysis, validity discussion

simple mean

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

Yes

Reasoning

Logical

null

['Procedurally-generated', 'LLM-generated']

['Targeted', 'Criterion']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['']

['No comparison made']

['Yes']

['Constructed']

['Mean']

luWebLINXRealworldWebsite2024

WEBLINX: Real-World Website Navigation with Multi-Turn Dialogue

Include

null

WebLinx introduces a large-scale benchmark of 100K interactions across 2300 expert demonstrations of conversational web navigation. Authors develop a multimodal agent able of interpreting both visual and textual input to complete web-based tasks with long context understanding and planning capabilities.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Multimodal reasoning, web navigation

Yes

“We define the real-world problem of conversational web navigation: given the initial user instruction, an agent must complete a real-world task inside a web browser while communicating with the user via multi-turn dialogue.” (Page 1, Introduction)

Subset

null

The task is defined as conversational web navigation, where an agent must complete a user-specified goal on a real-world website by interacting with the web interface (e.g., clicking, typing, submitting forms) while engaging in multi-turn dialogue with the user. The agent receives inputs such as browser screenshots, DOM elements, and dialogue history to predict the next action at each turn.

Each item is one step in a task, where the agent sees the current web page, past actions, and what the user said, and must decide what to do next—like clicking a button or typing text. Many of these steps together make up a full task.

Each task unfolds as a multi-turn dialogue between a user (called in the paper instructor) and an agent (called in the paper navigator), with actions done in a real browser environment. The task goal may not be fully known at the start and often evolves over the conversation, making long-term memory and contextual understanding important.

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Expert-crafted task examples (e.g. hand-written examples), Procedurally-generated task examples (e.g. Creating instances from a template)

"100K interactions across 2300 expert demonstrations of conversational web navigation."

Yes

website category, subcategory, geographic region, instructor visual access, AI assistance

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number), Extended interaction (e.g. conversation, calling an API and processing the response), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

The exact match metric is calculated for each turn in the task, comparing the predicted action (such as click, say, text input) with the ground truth action.

null

Academia

Yes

null

Test, Train, Validation

Train - 24418, validation - 1717

null

Simple Mean, aggregated using the micro-average of turn-level scores

Yes

Subscores are provided for element-based actions, text-based actions, and intent matching (whether the correct action type is predicted).

null

https://mcgill-nlp.github.io/weblinx/

WEBLINX

Contested

Yes

No

The benchmark is itself realistic

No

null

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

Benchmark is based on interacting with people, where the end can have different results, depending on how the conversation goes - therefore making it a bit of grey area.

Composite phenomenon

No

It is quite confusing from the text to count the size of the set, especially that it is based on interactions which are dependent variable - I am not fully sure if the sizes in task_dataset_size_extra are correct, I based them on table 8 as it is the only table showing active turns.

No

Agents

Web

null

['Real task', 'Author-crafted', 'Expert-crafted', 'Procedurally-generated']

['Targeted', 'Criterion']

['Short free response', 'Interaction', 'Structured']

['Exact match']

['Contested']

['Yes']

['Realistic']

['No']

['Complete']

null

zhangMuCGECMultireferenceMultisource2022

MuCGEC: a Multi-Reference Multi-Source Evaluation Dataset for Chinese Grammatical Error Correction

Include

null

The paper introduces MuCGEC, a multi-reference multi-source evaluation dataset for Chinese Grammatical Error Correction. This dataset contains different Chinese-as-a-Second-Language (CSL) learner sources, with each sentence corrected by three independent annotators and reviewed by a senior annotator.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Chinese Grammatical Error Correction

Yes

"Given a potentially noisy input sentence, grammatical error correction (GEC) aims to detect and correct all errors and produce a clean sentence"

Comprehensive

null

Models should detect and correct all grammatical errors in a given Chinese sentence, while preserving the original meaning.

Potentially erroneous Chinese sentence (input) paired with multiple human-annotated grammatically correct reference sentences (outputs).

null

Real task examples (e.g. GitHub issues)

1,092,285 (Lang8); 95,320 (HSK)

Yes

error types, sentence source, number of references per sentence, number of edits, character counts per sentence

Convenience sample (creators found a set of tasks that was readily accessible), Specific criteria (items were taken from a larger set based on specified rules)

Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train, Validation

1,996 sentences for NLPCC18-test, 2,000 for CGED-test, and 1,942 for Lang8-test

null

Simple Mean

Yes

(1) different data sources, (2) different error types, and (3) different numbers of references

null

https://github.com/HillZhang1999/MuCGEC

MuCGEC

Widely-agreed

Yes

No

Yes

1) establishing human performance baselines, (2) annotator error patterns, (3) demonstrating multiple references improve evaluation accuracy, and (4) demonstrate their character-based metrics.

Simple mean, char-based F0.5 scores for overall performance, along with precision and recall.

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

The task uses real sentences written by CSL learners that contain actual grammatical errors.

Composite phenomenon

Yes

null

No

NLP

Understanding

null

['Real task']

['Convenience', 'Criterion']

['Free response']

['Exact match', 'Soft match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Partial']

['Mean']

spragueMuSRTestingLimits2024

MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning

Include

null

The paper introduces MuSR, a long‑form natural‑language problems that require multistep “soft” reasoning combining commonsense, deductive and theory‑of‑mind inference. Generated through a neurosymbolic pipeline, these long-form problems reveal critical gaps in current models' reasoning capabilities despite being consistently solvable by humans.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

reasoning, commonsense knowledge, multi-step deductive

Yes

“Multistep soft reasoning” = reasoning that integrates narrative facts with implicit commonsense rules over several inference steps to reach an answer. First, a number of prior benchmarks do not have natural text. Others do not blend commonsense and multistep reasoning. Finally, we want a dataset that contains ground-truth intermediate structure and which is not solvable with rules. (section2)

Subset

null

Given a narrative and a question (multiple‑choice), the model is expected to choose the correct answer by reconstructing hidden reasoning chains.

a narrative (~1 000 words), a question, and answer options

Each item links to an underlying reasoning tree of ~10 steps and 6‑9 implicit commonsense facts.

Procedurally-generated task examples (e.g. Creating instances from a template), LLM-generated task examples (e.g. Filtered from responses to a prompt)

756

Yes

logic depth, number of commonsense facts, ground-truth intermediate facts

Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Instances produced by GPT‑4 via a neurosymbolic tree‑construction and “chaptering” pipeline.

Academia

No, link is broken

null

Test

null

Simple Mean

Yes

domain types, prompt variants, model baselines, prompt conditions

null

https://github.com/Zayne%E2%80%91Sprague/MuSR

MuSR

Contested

Yes

No

Yes

human/rule baselines, ablations, dataset difficulties

mean + standard deviatin, significance test. proportion

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

Synthetic but linguistically natural expressions, not a real‑world workflow or scenarios.

Single cohesive phenomenon

Not applicable

null

Yes

Reasoning

Commonsense

null

['Procedurally-generated', 'LLM-generated']

['Targeted']

['Multiple choice']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean', 'Std']

caoWenMindComprehensiveBenchmark2024

WenMind: A Comprehensive Benchmark for Evaluating Large Language Models in Chinese Classical Literature and Language Arts

Include

null

WenMind is a benchmark for Chinese Classical Literature and Language Arts (CCLLA). It spans 42 tasks across three sub-domains (Ancient Prose, Ancient Poetry, Ancient Literary Culture), in both domain- and capability-oriented formats.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Understanding, generation, and knowledge of Chinese classical language and literature

Yes

Understanding, generating, and applying knowledge of ancient Chinese texts across prose, poetry, and literary culture

Comprehensive

null

42 distinct tasks derived from classical Chinese language skills, such as translation, comprehension, poetry writing, idiom interpretation, etc.

Each item is a QA pair with metadata, eg. "Translate the following ancient Chinese sentence into modern Chinese."

Tasks categorized by sub-domain and cognitive ability (understanding, generation, knowledge).

Human exam questions (e.g. GRE questions), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

4875

Yes

Domain, capability, question format, task name (coarse/fine-grained), question, and answer

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics)

null

Internet exam databases, open-source Chinese text corpora (C2MChn, WYWEB), and LLM generations (ERNIE-3.5), all standardized and filtered into Q&A formats

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

Scores per task, domain (Prose/Poetry/Culture), and capability (Understanding/Generation/Knowledge)

null

https://github.com/SCUT-DLVCLab/WenMind

WenMind

Contested

Yes

No

Yes

Authors report 89.4% agreement between ERNIE-3.5 scoring and human evaluation across representative models and tasks

Stratified human agreement evaluation on LLM-graded items; comparisons to BLEU/F1 for scoring validity.

Outputs alone

Representative task (e.g. answering medical licensing exam questions), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

Maybe representative of how LLMs may be used in educational and cultural settings, but mostly a "knowledge test".

Composite phenomenon

Yes

null

No

Knowledge

Cultural

null

['Human exams', 'Author-crafted', 'Another benchmark', 'LLM-generated']

['Convenience', 'Targeted', 'Criterion']

['Multiple choice', 'Short free response', 'Free response']

['Exact match', 'Soft match', 'LLM-as-a-Judge']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Representative', 'Constructed']

['Mean']

ouyangCliMedBenchLargeScaleChinese2024

CliMedBench: A Large-Scale Chinese Benchmark for Evaluating Medical Large Language Models in Clinical Scenarios

Include

null

CliMedBench is a Chinese clinical medical benchmark with 33.7k QA items across 14 core scenarios derived from real-world medical records and exams, measuring LLMs’ clinical reasoning and language abilities. It includes evaluations of 11 LLMs.

Proposes a novel adaptive testing method (agent-based CAT) grounded in Item Response Theory to alleviate cost concerns, but stresses the "real" benchmark is still just evaluating all models across all questions

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

medical reasoning, diagnostic accuracy, clinical QA

Yes

Defines clinical reasoning and related capabilities such as hallucination resistance, information retrieval, and instruction via the "personas" of real medical practitioners eg. radiographer; pharmacist

Subset

The selected "personas" feel a bit random: doctor, med student, patient, radiologist, pharmacist. Surely there are other people in a hospital.

Multiple-choice, sequencing, and open-ended questions based on real-world Chinese clinical scenarios.

One task item can be a MCQ from an EHR, a sequencing task like reordering surgical steps, or open-ended responses like writing discharge summaries

null

Human exam questions (e.g. GRE questions), Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), LLM-generated task examples (e.g. Filtered from responses to a prompt)

33735

Yes

clinical role, task type, scenario ID, source (e.g., EHR, exam), evaluation axes if relevant

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Free response (e.g. summary paragraph, executable code), Sequencing

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF), Human ratings (text quality, preference, NOT manual scoring of other metrics)

Mix of automatic metrics and expert review

Combination of electronic health records, exam data, expert rephrasing, and LLM generation with filtering

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

14 clinical scenarios and 7 evaluation axes

null

https://github.com/Optifine-TAT/CliMedBench

CliMedBench

Contested

Yes

No

Yes

Medical professionals rated benchmark scenarios and Spearman correlation to another benchmark (MedBench) was computed

Simple mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Representative task (e.g. answering medical licensing exam questions)

Evaluation of actual clinical decisions/tasks, but in a segmented way

Composite phenomenon

Yes

null

No

Medicine

null

['Human exams', 'Real task', 'Author-crafted', 'LLM-generated']

['Targeted', 'Criterion']

['Multiple choice', 'Free response', 'Free response']

['Exact match', 'Soft match', 'Human ratings']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Partial', 'Representative']

['Mean']

sabourEmoBenchEvaluatingEmotional2024

EmoBench: Evaluating the Emotional Intelligence of Large Language Models

Include

null

EmoBench evaluates emotional intelligence (EI) covering emotional understanding and application with 400 questions in English and Chinese.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Emotional Intelligence

Yes

Emotional Intelligence, ""the ability to monitor feelings of our own and understand feelings of others, differentiate between them, and leverage this information to guide our thoughts and actions"". Authors use understanding and application as core components.

Comprehensive

Use both a breakdown of possible emotions and a breakdown into application and understanding

Scenario-based multiple-choice questions; some just requiring selection of most effective action, others requiring action selection + definition of emotion in scenario

Scenario + MCQ on emotions, causes, or best action

Task items are theory-grounded and include taxonomies of emotions

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

400

Yes

Emotion type, correct label, category

Unknown

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

GPT-4 used for inspiration in creating scenarios, final content human-authored

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

By task type, language, subcategory

null

https://github.com/Sahandfer/EmoBench

EmoBench

Contested

Yes

No

Yes

Exhaustive human baselining via online survey of 48 participants (humans outperform all LLMs on average)

Accuracy, Fleiss’ k for human agreement

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

Fully fictional; while grounding in psychological theory seems very solid, task is completely from-scratch, not e.g. from existing personality/emotional intelligence exams (should those exist)

Composite phenomenon

Yes

null

Yes

Psychology

null

['Author-crafted']

['Unknown']

['Multiple choice']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean', 'Other']

garcia-ferreroThisNotDataset2023

This is not a Dataset: A Large Negation Benchmark to Challenge Large Language Models

Include

null

The paper introduces a large semi-automatically generated dataset of circa 400,000 descriptive sentences about commonsense knowledge that can be true or false in which negation is present in about 2/3 of the corpus in different forms.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

negation probing

Yes

Although large language models (LLMs) have apparently acquired a certain level of grammatical knowledge and the ability to make generalizations, they fail to interpret negation, a crucial step in Natural Language Processing.

Subset

null

Take a natural language sentence, and classify its truth value.

A brother is never a female person who has the same parents as another person. True.

null

Expert-crafted task examples (e.g. hand-written examples)

90,281

Yes

types and amount of negative knowledge

Targeted items (creators defined a task space and chose tasks within it strategically)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test, Train, Validation

Train: 268,505; validation: 2,514

null

Yes

By types and amount of negative knowledge

null

https://github.com/hitz-zentroa/This-is-not-a-Dataset

null

Widely-agreed

Yes

No

No comparisons made

No

Yes

3.2 Dataset quality assessment: Human Evaluation addresses the validation of the generation process and the different templates used, that is to say, whether the sentences in the dataset are grammatical and that overall represent true and false knowledge as expected.

null

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

No

NLP

Understanding

null

['Expert-crafted']

['Targeted']

['Short free response']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Constructed']

null

josephFactPICOFactualityEvaluation2024

FACTPICO: Factuality Evaluation for Plain Language Summarization of Medical Evidence

Include

null

FACTPICO evaluates the factuality of LLM-generated plain language summaries of medical randomized controlled trials (RCTs). It features fine-grained expert annotations across five key dimensions and includes both human and LLM-generated rationales.

Rich analysis, but incredibly unscalable - all evaluation was done by expert humans, which explains why they only tested 3 models

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Factuality in plain language summarization

Yes

Factuality is defined as accurate representation of critical trial elements (Population, Intervention, Comparator, Outcome - that's PICO) and their results, with particular focus on correctness of added explanatory content

Subset

null

Generate plain-language summaries of abstracts of randomized controlled trials.

Each item is an RCT abstract, the output always its summary.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language)

115

Yes

PICO annotations, human rationales.

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Free response (e.g. summary paragraph, executable code)

Human ratings (text quality, preference, NOT manual scoring of other metrics)

Time-intensive human scoring of free-text responses across the PICO dimensions

Based on RCT descriptions from Evidence Inference 2.0 dataset, sampled for exclusion of those which already have human summaries

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

Subscores for each PICO element and Evidence Inference

null

https://github.com/lilywchen/FactPICO

FACTPICO

Contested

Yes

No

No comparisons made

The benchmark is itself realistic

Yes

Authors analyze inter-annotator agreement, correlation with expert judgments, and rationale similarity.

Flesch Kincaid, Rouge-L, Kendalls, Spearmans

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

Summarizing technical trial outcomes into accessible language is a real-world medical communication need.

Single cohesive phenomenon

Not applicable

null

No

NLP

Summarization

null

['Author-crafted', 'Another benchmark']

['Targeted', 'Criterion']

['Free response']

['Human ratings']

['Contested']

['Yes']

['Realistic']

['Yes']

['Partial']

['Other']

liuRevisitingDeIdentificationElectronic2023

Revisiting De-Identification of Electronic Medical Records: Evaluation of Within- and Cross-Hospital Generalization

Include

null

Boundary include - measures LLM capabilities in theory and does provide a benchmark, but tests pre-ChatGPT models including a self-trained CNN

Benchmark for de-identification of protected health information (PHI) in Chinese electronic medical records, with a focus on cross-hospital generalization. Constructs a multi-hospital dataset and evaluates various models and domain generalization (DG) techniques to assess performance under domain shift.

Pre-LLMs so uses the dataset - which is still a valid benchmark - as training data for a CNN and a BERT fine-tuning run.

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Generalization on de-identification tasks

Yes

Anonymization

Subset

null

Detect and remove personal health information mentions (e.g., names, locations, dates) in clinical records from three Chinese hospitals.

Each item is a sentence or span from an electronic medical record, with relevant tokens labeled using tags corresponding to personal data categories, but the labels hidden. The task is to recreate these token labels.

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

500

Yes

Sentence and mention counts, health information category counts per dataset

Random sample (creators defined a task space and sampled from it), Convenience sample (creators found a set of tasks that was readily accessible)

Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

Predicted span of tags must exactly match correct span and category

Data collected from three Chinese hospitals - no idea how that got around data protection laws even with anonymization; they must be quite loose there - and hand-annotated

Academia

Yes

null

Test, Train, Validation

400

Sequence output with personal data tagged

Simple Mean

Yes

Per PII category (e.g., PERSON, DATE, ID)

null

https://github.com/lanyangyang93/Revisiting-De-Identification

null

Widely-agreed

Yes

No

No comparisons made

The benchmark is itself realistic

No

null

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

De-identification is a required step for real-world medical data sharing and lots of other data sharing contexts.

Single cohesive phenomenon

Not applicable

null

No

Medicine

null

['Real task', 'Author-crafted']

['Random', 'Convenience']

['Structured']

['Exact match']

['Widely-agreed']

['Yes']

['Realistic']

['No']

['Complete']

null

liuConvBenchMultiturnConversation2024

ConvBench: A Multi-Turn Conversation Evaluation Benchmark with Hierarchical Ablation Capability for Large Vision-Language Models

Include

null

This paper introduces the ConvBench to evaluate LVLMs across hierarchical capabilities such as perception, reasoning, and creativity. It enables fine-grained error attribution and includes an automatic evaluation framework.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

multi-turn visual conversation

Yes

Open-ended multi-turn visual conversations

Comprehensive

null

Take an image, and provide answer during multi-turn conversation

<image> user question 1, model answer 1, user question 2, model answer 2

null

Expert-crafted task examples (e.g. hand-written examples), Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

577

Yes

hierarchy of multimodal capabilities

Specific criteria (items were taken from a larger set based on specified rules)

Free response (e.g. summary paragraph, executable code)

LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics)

null

Academia

Yes

null

Test

null

Simple Mean, Weighted Mean

No

null

https://github.com/shirlyliu64/ConvBench

ConvBench

Widely-agreed

Yes

No

No comparisons made

No

Yes

No

null

simple mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

No

User Interaction

null

['Expert-crafted', 'Another benchmark', 'LLM-generated']

['Criterion']

['Free response']

['LLM-as-a-Judge']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Partial', 'Representative']

['Mean']

dinhSciExBenchmarkingLarge2024

SciEx: Benchmarking Large Language Models on Scientific Exams with Human Expert Grading and Automatic Grading

Include

null

SciEx is a multilingual, multimodal benchmark of university-level computer science exams. It includes freeform questions with expert grading that contain both text and images. It compares to a baseline of human students, as all questions are real exam questions.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

scientific reasoning, problem solving

Yes

"Solve scientific tasks" via exam-style problem solving, including reasoning, proof generation, and algorithmic thinking

Subset

null

Answer university CS exam questions, some of which are multimodal and freeform, with human-graded performance.

Each item consists of a CS exam question (text and possibly images), and the LLM must generate a free-text or structured response.

Questions span various formats, languages (English and German), and topics (e.g. AI, databases, algorithms)

Human exam questions (e.g. GRE questions)

154

Yes

max score, student average, gold reference answer, difficulty level, language, modality

Convenience sample (creators found a set of tasks that was readily accessible)

Free response (e.g. summary paragraph, executable code), Structured response (e.g. valid JSON, API call alone)

Human ratings (text quality, preference, NOT manual scoring of other metrics)

Expert graders score each answer; LLM-as-a-judge methods are evaluated via correlation with these expert scores but not used in the benchmark

CS exams from Karlsruhe Institute of Technology (2022–2024), authored by instructors

Academia

Yes

null

Test

null

Simple Mean

Yes

Grouped by difficulty, modality, and language

null

https://github.com/TuAnh23/SciEx

SciEx

Contested

Yes

No

Yes

Comparison of expert vs LLM grading; examined factors like image use, question difficulty, language; measured grader bias. Detailed comparison to student performance, though number of students is not listed.

Pearson correlation, RMSE, differences to student baselines

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

Reflects typical LLM use cases in academic or tutoring settings, but the authors specifically claim the benchmark is for scientific work, which makes QA on exam questions representative at best

Composite phenomenon

No

null

No

General Science

null

['Human exams']

['Convenience']

['Free response', 'Structured']

['Human ratings']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Other']

zhangUnveilingTapestryConsistency2024

Unveiling the Tapestry of Consistency in Large Vision-Language Models

Include

null

This paper introduces the ConBench benchmark to evaluate the consistency of LVLMs across prompts with varying solution spaces centered on the same knowledge point. The authors reveal key patterns in LVLM behavior and propose a trigger-based diagnostic refinement method to improve consistency and indirectly enhance captioning performance.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

consistency

Yes

Although LVLMs can generate high-quality responses to task prompts, we discover that for correctly answered cases, simply modifying the prompt will result LVLMs in providing contradictory responses.

Comprehensive

null

Take an image and question, provide a short-form answer

<image> Question: How many real cats in the image. A) One B) Two C) Three D) Four. Answer: A

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language)

4000

Yes

Hierarchical Core Capability, Question Type

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

Hierarchical Core Capability, Question Type

null

https://github.com/foundation-multimodal-models/ConBench

null

Widely-agreed

Yes

No

No comparisons made

No

null

simple mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Representative task (e.g. answering medical licensing exam questions)

null

Single cohesive phenomenon

Not applicable

null

No

Language Modelling

Robustness

null

['Author-crafted', 'Another benchmark']

['Targeted', 'Criterion']

['Multiple choice', 'Short free response']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Partial', 'Representative']

['Mean']

fierroMuLanStudyFact2024

MULAN : A Study of Fact Mutability in Language Models

Include

null

The authors create MULAN , a benchmark for evaluating the ability of English language models to anticipate time-contingency, covering both 1:1 and 1:N relations.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

anticipating when facts are time-contingent

Yes

time awareness in LLMs, specifically for encoding of fact mutability in their representations and for the comparative ease of editing of mutable facts versus immutable ones.

Subset

null

Given a subject–relation query (input), the task is to predict the correct object(s) (output), where queries may involve either immutable or mutable facts.

subject–relation query (input), correct object(s) (output).

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

Test: 35000 / 31410.

Yes

Immutable or not.

Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train, Validation

Train: 6230 / 6820; Validation: 5780 / 5910.

null

Yes

Immutability-1 and Immutability-N.

null

https://github.com/coastalcph/fact_mutability

MULAN

Contested

Yes

null

No

No comparisons made

No

null

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

No

null

No

Reasoning

Temporal

null

['Author-crafted']

['Criterion']

['Short free response']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['No']

['Constructed']

null

bitton-guettaVisualRiddlesCommonsense2024

Visual Riddles: a Commonsense and World Knowledge Challenge for Large Vision and Language Models

Include

null

This paper introduces Visual Riddles, a benchmark designed to evaluate LLMs on complex visual reasoning tasks that require commonsense and world knowledge. The dataset includes 400 carefully crafted riddles, each combining images, questions, and textual hints, revealing significant performance gaps between current models and human reasoning abilities.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

knowledge, visual understanding

Yes

While humans easily recognize such contextual nuances, existing image-understanding models struggle to integrate visual cues with world knowledge stemming from cultural aspects, life-experiences, and physical or social knowledge

Comprehensive

null

(1) Main Task: Solve open-ended questions. (2) Utilizing Hints: Use textual aids to identify key visual clues in riddles. (3) Employing Attributions: Apply web-sourced attributions to improve world-knowledge. (4) Multiple Choice: Select the correct answer to the riddle from five options. (5) Automatic Evaluation: Evaluate open-ended answers in two scenarios— Reference-Free, assessing the correctness of a candidate answer (CA) based only on the visual riddle, and Reference-Based, comparing CAs to the ground truth answer (GTA).

null

it covers five different subtasks

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Expert-crafted task examples (e.g. hand-written examples)

500

No

null

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), Human ratings (text quality, preference, NOT manual scoring of other metrics), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics)

They use human rating as primary evaluation method, while also compare the performance between LLM-as-a-judge and human rating in ablation study

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Yes

Open-ended VQA, Multiple-choice VQA, Open-ended VQA Automatic Evaluation

null

https://huggingface.co/datasets/visual-riddles/visual_riddles

Visual Riddles

Widely-agreed

Yes

No

No comparisons made

No

null

simple mean

Outputs alone

Representative task (e.g. answering medical licensing exam questions), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

No

VQA

null

['Author-crafted', 'Expert-crafted']

['Targeted', 'Criterion']

['Multiple choice', 'Short free response', 'Free response']

['Exact match', 'Human ratings', 'LLM-as-a-Judge']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Representative', 'Constructed']

['Mean']

liQuantifyingAdaptabilityPretrained2022

Quantifying Adaptability in Pre-trained Language Models with 500 Tasks

Include

null

TaskBench500 is a benchmark designed to systematically measure how LLMs adapt to new tasks. It comprises "500 procedurally generated sequence modeling tasks" spanning "lexical semantics, sequence processing, memorization, logical reasoning, and world knowledge" (4696). The benchmark is publicly available.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

adaptability, compositional adaptability

Yes

Language model (LM) adaptability is how LMs are tuned to "perform" a "new task" they were not trained to complete, with adaption techniques like "fine-tuning or prompt-tuning" (4696).

Comprehensive

Task adaption is split into the domains of measuring memory, composition, and distribution matching.

The benchmark defines 500 tasks total tasks. It first defines atomic tasks that are then "combined using a set of composition operators to produce more complex tasks" (4698). Atomic tasks span lexical tasks, factual tasks, and random relation tasks, and composition operators include word-level and sequential compositions.

Although there are 500 distinct tasks, "every task takes as input a word or word sequence, and outputs either a boolean value or a set of words/word sequences" (4698).

null

Modified from another benchmark (e.g. translation into another language), Procedurally-generated task examples (e.g. Creating instances from a template)

null

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), Distribution (perplexity, calibration, correlation), Correlation (Matthew's correlation, Pearson's r)

To measure memorization, accuracy is used. To measure compositions of atomic tasks, Pearson correlation is used. Both metrics are referred to as the adaptability metric for their task. To measure how models learn new distributions, the paper defines a custom metric to produce "an aggregated probability mass assigned to all easier task and all harder tasks in a task pair," so it should be easier to adapt to an easier task than a harder task (4703).

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

Atomic, Word-level Comp, Seq Comp

null

https://github.com/facebookresearch/task_bench

TaskBench500

Contested

Yes

No

No comparisons made

No

Yes

The authors distinguish between previous attempts to study "generalization to new examples" and the paper's "systematic study of adaption to new tasks" (4698). The authors identify that "new pre-raining adaption schemes are evaluated using small suites of curated tasks" which are "poorly suited for answering larger, structural questions" like "can we predict how quickly (and how effectively) pre-trained LMs can be adapted to perform it" (4696).

Simple mean, average

Outputs alone

Representative task (e.g. answering medical licensing exam questions), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

No

null

No

Language Modelling

In-context Learning

null

['Another benchmark', 'Procedurally-generated']

['Convenience', 'Targeted', 'Criterion']

['Short free response', 'Free response']

['Exact match', 'Distribution', 'Correlation']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Representative', 'Constructed']

['Mean']

duEmbSpatialbenchBenchmarkingSpatial2024

EmbSpatial-Bench: Benchmarking Spatial Understanding for Embodied Tasks with Large Vision-Language Models

Include

null

This paper introduces the EmbSpatial-Bench benchmark to evaluate the spatial understanding capabilities of LVLMs in embodied environments. The authors also propose EmbSpatial-SFT, an instruction-tuning dataset aimed at enhancing LVLMs' spatial reasoning from an egocentric perspective

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

spatial understanding

Yes

understand spatial relationships between object in embodied scenarios

Subset

null

Take a spatial image and multiple-choice question, provide the answer

<spatial image> Question: How are television and shelf positioned in relation to each other in the image? A: ..., B: ..., C: ..., D: ....

null

Procedurally-generated task examples (e.g. Creating instances from a template), LLM-generated task examples (e.g. Filtered from responses to a prompt)

3640

No

null

Random sample (creators defined a task space and sampled from it)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

For likelihood strategy, which uses the option with the highest probability generated by the model, the model output logits is needed

null

Academia

Yes

null

Test

null

Simple Mean

No

null

https://github.com/mengfeidu/EmbSpatial-Bench

EmbSpatial-Bench

Widely-agreed

Yes

No

No comparisons made

No

Yes

No

null

simple mean

Model access required (e.g. logits)

Representative task (e.g. answering medical licensing exam questions), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

No

VQA

null

['Procedurally-generated', 'LLM-generated']

['Random']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Representative', 'Constructed']

['Mean']

chungCanVisualLanguage2024

Can visual language models resolve textual ambiguity with visual cues? Let visual puns tell you!

Include

null

This paper introduces the UNPIE benchmark to evaluate multimodal understanding in machines using puns. By pairing puns with explanatory images, the study tests models on tasks like grounding, disambiguation, and reconstruction, showing that visual context significantly enhances performance over text-only approaches.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

visual understanding

Yes

Humans possess multimodal literacy, allowing them to actively integrate information from various modalities to form reasoning. Faced with challenges like lexical ambiguity in text, we supplement this with other modalities, such as thumbnail images or textbook illustrations. Is it possible for machines to achieve a similar multimodal understanding capability?

Comprehensive

null

(1) pun grounding: to identify the specific phrase in a sentence that forms a pun (2) pun disambiguation: to choose the translation that best matches the image provided as a pun disambiguator. (3) pun reconstruction: to recreate the original English pun sentence using a translated version with potentially no ambiguity.

null

Three different tasks are proposed

Expert-crafted task examples (e.g. hand-written examples), LLM-generated task examples (e.g. Filtered from responses to a prompt)

1000

Yes

language

Specific criteria (items were taken from a larger set based on specified rules)

image

Exact Match (accuracy, F1, precision, recall)

null

To generate pun explanation images, the expert annotators prompt the DALLE 3 model to create images; The paper also designs a cooperative framework between machines and humans for pun translation

Academia

Yes

null

Test

null

Simple Mean

Yes

The report scores for different tasks (subsets) separately

null

https://github.com/JiwanChung/VisualPun_UNPIE

UNPIE

Widely-agreed

Yes

No

No comparisons made

No

null

simple mean/sum

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

No

Grounding

null

['Expert-crafted', 'LLM-generated']

['Criterion']

['Free response']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Constructed']

['Mean']

kumarVisionlanguageModelsUnderstand2024

Do Vision-Language Models Understand Compound Nouns?

Include

null

Compun is a multimodal benchmark to assess how models understand compound nouns using text-to-image retrieval tasks. The dataset is publicly available, manually curated, and focuses on noun+noun compound nouns.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

compound nouns

Yes

"A compound noun (CN) is a noun formed from two or more words combined to create a single noun with a new meaning." (519).

Subset

Compun focuses "primarily on the noun + noun type" of compound nouns (519).

The task is "text-to-image retrieval where, given a text prompt with a CN [compound noun]" the model must "select the correct image that shows the CN among a pair of distractor images that show the constituent nouns that make up the CN" (519).

"Each instance in Compun corresponds to a unique compound noun and includes one image representing the compound noun, along with two additional distractor images" (520).

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language), Procedurally-generated task examples (e.g. Creating instances from a template), LLM-generated task examples (e.g. Filtered from responses to a prompt)

400

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

The paper defines a binary accuracy metric, where the score is 1 if the cosine similarity of the prompt and positive caption is greater than the similarities between the prompt and the negative captions, and 0 otherwise.

The paper defines a new pipeline to generate prompts for text-to-image retrieval, where given a compound noun, a LLM generates "multiple diverse captions" where "each caption describes a scene with the compound noun as a key object in it. Finally, the captions are used to construct a custom prompt for text-to-image retrieval" (520). Expert annotators manually collect the images per instance, and MTurk was used to decide the most common meaning if a compound noun had several interpretations.

Academia

Yes

null

Test

null

Simple Mean

No

null

https://github.com/sonalkum/Compun

Compun

Widely-agreed

Yes

No

No comparisons made

No

Yes

The authors ground the text-to-image retrieval task in previous cognitive science research, and discuss how the benchmark can be improved by expanding to different types fo compound nouns and by using novel metrics for retrieval.

Average

Model access required (e.g. logits)

Representative task (e.g. answering medical licensing exam questions), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

No

null

No

NLP

Understanding

null

['Author-crafted', 'Another benchmark', 'Procedurally-generated', 'LLM-generated']

['Convenience', 'Criterion']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative', 'Constructed']

['Mean']

flachsGrammaticalErrorCorrection2020

Grammatical Error Correction in Low Error Density Domains: A New Benchmark and Analyses

Include

null

CWEB is a benchmark for grammatical error correction that is publicly available and manually annotated by experts. It contains website data from Common Crawl, and includes sentences with low and high error density.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Grammatical error correction

Yes

"Grammatical error correction (GEC) is the task of automatically editing text to remove grammatical errors" (8467).

Comprehensive

null

The model is given text and must identify and correct grammatical errors.

A single item contains a sentence with in-line corrections.

null

Modified from another benchmark (e.g. translation into another language)

13574

No

null

Random sample (creators defined a task space and sampled from it)

Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), Distribution (perplexity, calibration, correlation)

The paper uses F_{0.5} and ERRANT, standard grammar error correction metrics, cited on page 8471, to assess the correctness of the correction. Perplexity and semantic similarity are used to measure the semantic change in a sentence after the edit.

The websites are in English and derived from the first 18 dumps of Common Crawl. Text is filtered to remove non-English and incomplete sentences using justText. The data is manually corrected by expert annotators. The dataset is split into CWEB-S (sponsored websites) and CWEB-G (generic) websites.

Mix (multiple authors from industry and academia)

Yes

null

Test, Train

Development/Test 6729/6845

null

No

null

https://github.com/SimonHFL/CWEB

CWEB (Corrected WEBsites)

Widely-agreed

Yes

The benchmark is itself realistic

No

Yes

The authors detail that grammar error correction (GEC) models must "perform well in the open-domain setting and generalize, not only to writing produced in the educational context, but also to language production 'in the wild'" (8647). The authors also highlight that a strong GEC benchmark must evaluate "domain adaptation and low precision" in texts with low error density (8647).

Simple mean

Model access required (e.g. logits)

Complete real task (e.g. providing medical advice to real people interactively)

null

Single cohesive phenomenon

No

null

No

NLP

Understanding

null

['Another benchmark']

['Random']

['Free response']

['Exact match', 'Distribution']

['Widely-agreed']

['Yes']

['Realistic']

['Yes']

['Complete']

['Mean']

joshiILTURBenchmarkIndian2024

IL-TUR: Benchmark for Indian Legal Text Understanding and Reasoning

Include

null

This paper introduces IL-TUR, a benchmark designed to evaluate NLP models for legal text understanding and reasoning in the Indian legal context. The benchmark covers both monolingual (English, Hindi) and multilingual tasks across 9 Indian languages. It provides baseline models and a leaderboard with comparison in automating legal document processing.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Legal text understanding and reasoning in Indian and multilingual context

Yes

The tasks should cater exclusively to the legal domain. Solving a task should require in-depth knowledge and understanding of the law and its associated areas. [...] Moreover, solving legal tasks should require knowledge about the law as well as commonsense knowledge and societal norms about the world.

null

There are 6 types of tasks: sequence classification, multi-class classification, Classification and extraction, classification, multi-label classification, retrieval, and generation

text (text, short text, paragraph, sentences) and the output label or text

null

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Expert-crafted task examples (e.g. hand-written examples), Crowd-sourced task examples (e.g. Prolific-created tasks), Procedurally-generated task examples (e.g. Creating instances from a template)

~15,100, but the information is provided for some tasks and datasets

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF), BERTScore, GLEU

null

Academia

Yes

null

Test, Train, Validation

some tasks/datasets have this info some not

null

Simple Mean

No

null

majority@k (majority vote over k trials)

https://exploration-lab.github.io/IL-TUR/

IL-TUR

Contested

Yes

No

No comparisons made

No

Yes

"The tasks should cater exclusively to the legal domain. Solving a task should require in-depth knowledge and understanding of the law and its associated areas [...] as well as commonsense knowledge and societal norms."

simple mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Representative task (e.g. answering medical licensing exam questions), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

CJPE (Court Judgment Prediction with Explanation) or BAIL (Bail Prediction) are framed as supervised classification tasks using labeled court documents, which is a constructed task. Legal Named Entity Recognition (L-NER) and Rhetorical Role Prediction (RR), which model foundational capabilities for downstream legal NLP applications are representative tasks to model the legal system, the rest are partial real tasks as the use case is the real legal docs

Composite phenomenon

Yes

null

No

Law

null

['Real task', 'Author-crafted', 'Expert-crafted', 'Crowd-sourced', 'Procedurally-generated']

['Convenience', 'Targeted', 'Criterion']

['Short free response', 'Free response']

['Exact match', 'Soft match', 'Soft match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Partial', 'Representative', 'Constructed']

['Mean']

royBenchCLAMPBenchmarkEvaluating2023

BenchCLAMP: A Benchmark for Evaluating Language Models on Syntactic and Semantic Parsing

Include

null

BenchCLAMP evaluates Constrained LAnguage Model Parsing on syntactic and semantic parsing tasks. It provides context-free grammars using both prompt-based learning and fine-tuning across different data resource settings.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

syntactic parsing, semantic parsing

Yes

"Parsing tasks are generally not considered a testbed for such evaluation. The outputs of parsing tasks are structured objects such as parse trees or code. State-of-the-art systems thus involve task- or dataset-specific model architectures and target representation constraints. Evaluating language models on parsing tasks test capabilities not captured by commonly used evaluation tasks."

Comprehensive

null

Generating syntactic or semantic representations from natural language inputs, which includes context-free grammars for semantic and syntactic parsing datasets, as well as a constrained decoding interface to generate only valid outputs covered by these grammars.

Natural language utterance; structured output representation (constituency parse tree, dependency relations, formal meaning representation)

null

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language)

train: 5500

Yes

dataset type, output formalism, evaluation metric

Convenience sample (creators found a set of tasks that was readily accessible), Specific criteria (items were taken from a larger set based on specified rules)

Free response (e.g. summary paragraph, executable code), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Industry

Yes

null

Test, Train

train:5500, dev: 550

null

Simple Mean

Yes

Data resource settings; parsing datasets; constraint settings

null

https://github.com/microsoft/semantic_parsing_with_constrained_lm

null

Widely-agreed

Yes

No

null

Mean scores across different data splits and standard deviation for low-resource settings

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Representative task (e.g. answering medical licensing exam questions)

While not all examples are derived from real user interactions, they are designed to represent realistic use cases for parsing.

Composite phenomenon

Yes

null

Yes

NLP

Understanding

null

['Real task', 'Another benchmark']

['Convenience', 'Criterion']

['Free response', 'Structured']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Partial', 'Representative']

['Mean', 'Std']

ryanRevisitingNonEnglishText2023

Revisiting non-English Text Simplification: A Unified Multilingual Benchmark

Include

null

MULTI-SIM, benchmark for multilingual text simplification, containing complex-simple sentence pairs across 12 languages. They show improvements from multilingual training for non-English languages and strong performance of Russian for cross-lingual transfer.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

automatic text simplification, multilingual

Yes

"Automatic text simplification (ATS) is the task of reducing the complexity of a text without chang- ing its original content and meaning"

Comprehensive

null

Automatic Text Simplification transforms complex sentences into simpler versions that maintain the original meaning but reduce linguistic complexity.

Pair of sentences: a complex sentence and its corresponding simpler version that preserves the original meaning.

null

Real task examples (e.g. GitHub issues), Expert-crafted task examples (e.g. hand-written examples), Modified from another benchmark (e.g. translation into another language)

train: 653,468

Yes

language, script, domain, collection approach, simplification type

Convenience sample (creators found a set of tasks that was readily accessible), Specific criteria (items were taken from a larger set based on specified rules)

Free response (e.g. summary paragraph, executable code)

n-gram (BLEU, ROUGE, chrF), Human ratings (text quality, preference, NOT manual scoring of other metrics)

null

Academia

Yes

null

Test, Train

test: 6,306; dev: 6,728

null

Simple Mean

Yes

Scores for each language dataset separately, single-dataset fine-tuning, joint language fine-tuning, zero-shot cross-lingual transfer

null

https://github.com/XenonMolecule/MultiSim

MULTI-SIM

Contested

Yes

No

No comparisons made

Yes

(1) human evaluation to assess the quality of model outputs on their benchmark; (2) inter-annotator agreement using Krippendorff's alpha; (3) analyze corpus statistics to understand dataset quality; (4) acknowledge limitations "we cannot guarantee the quality of each resource or validate the methods that the original authors used to create them"

Automatic Evaluation Metrics (SARI; BLEU); Measure inter-annotator agreement using Krippendorff's alpha

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

Sentence pairs come from real-world simplification efforts intended for actual audiences with lower literacy levels.

Composite phenomenon

No

null

Yes

NLP

Summarization

null

['Real task', 'Expert-crafted', 'Another benchmark']

['Convenience', 'Criterion']

['Free response']

['Soft match', 'Human ratings']

['Contested']

['Yes']

['Comparison made']

['Yes']

['Partial']

['Mean']

siREADINChineseMultitask2023

READIN: A Chinese Multi-Task Benchmark with Realistic and Diverse Input Noises

Include

null

READIN: a Chinese multi-task benchmark with REalistic And Diverse Input Noises. READIN contains four diverse tasks and requests annotators to re-enter the original test data with two commonly used Chinese input methods: Pinyin input and speech input.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Robustness to realistic input noises

Yes

Robustness to "realistic and diverse input noises" in Chinese NLP, specifically focusing on user-generated inputs in real-world applications.

Subset

null

Comparing performance on clean data versus data with keyboard and speech input noises.

Clean test example from an existing Chinese NLP dataset paired with multiple noisy versions of the same example, along with the original expected output.

null

Crowd-sourced task examples (e.g. Prolific-created tasks), Modified from another benchmark (e.g. translation into another language)

train: 34,371

Yes

Input noise type, error rate, annotator information, original task type

Convenience sample (creators found a set of tasks that was readily accessible), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF)

null

Academia

Yes

null

Test, Train, Validation

dev: 4,430; test: 8,570

Response format varies depending on the task

Micro-Average, Worst-Average

Yes

input noise type (keyboard/speech)

null

https://github.com/thunlp/READIN

READIN

Widely-agreed

Yes

The benchmark is itself realistic

No

Yes

(1) Human evaluations showing plausibility of crowd-sourced noises; (2) Diversity analysis of collected noises; (3) Character-level error rates; quantify noise in test sets; (4) Qualitative case studies

Micro-Average, Worst-Average

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

Simulates real-world user input scenarios where people make typing errors or speech recognition systems misinterpret accented speech.

Composite phenomenon

Yes

null

No

Language Modelling

Robustness

null

['Crowd-sourced', 'Another benchmark']

['Convenience', 'Criterion']

['Multiple choice', 'Short free response', 'Free response', 'Structured']

['Exact match', 'Soft match']

['Widely-agreed']

['Yes']

['Realistic']

['Yes']

['Partial']

['Mean']

schwettmannFINDFunctionDescription2023

FIND: A Function Description Benchmark for Evaluating Interpretability Methods

Include

null

Boarderline. Evaluates LLM-based automated interp methods. Ultimately does evaluate capabilities of LLMs so including.

Evaluating the ability of LLMs as interpretability agents as a proxy for how well they might perform in automated interpretability pipelines. i.e. can LLMs recover the functions from input/output data alone.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Automated interpretability capabilities

No

null

Comprehensive

Very broad phenomena "evaluating automated interpretability methods"

Recover the details about a program (e.g. a maths program/string program...) from examples of just the input and outputs.

A ground truth function and an operator that might add some noise to the data. Operationalised in an agentic way where the model can call the function many times and examine its behaviour.

Very simplified version of the real-life phenomena.

Procedurally-generated task examples (e.g. Creating instances from a template)

null

Yes

Type of function (numeric, string, and synthetic neural modules)

Random sample (creators defined a task space and sampled from it), Targeted items (creators defined a task space and chose tasks within it strategically)

Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics), Execution-based (unit tests)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

Function type (numeric, string, neural modules)

null

https://github.com/multimodal-interpretability/FIND

FIND

Contested

Highly simplified version of the phenomena

Yes

No

No comparisons made

No

Yes

Discuss that the benchmark is a test of "necessary, but not sufficient, capabilities for automated interpretation." "The ultimate test of these interpretation methods’ effectiveness must be their ability to generate actionable insights about real models, which FIND does not evaluate."

Mean

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

2,275

Single cohesive phenomenon

Not applicable

null

No

Reasoning

null

['Procedurally-generated']

['Random', 'Targeted']

['Free response']

['Exact match', 'LLM-as-a-Judge', 'Reward']

['Contested']

['Partially']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean']

zuoPatentEvalUnderstandingErrors2024

PatentEval: Understanding Errors in Patent Generation

Include

null

PatentEval, a benchmark annotated by human experts, tailored for assessing language models of different sizes and capacities. This includes pairwise comparisons and detailed analysis of error types in each output.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

patent text evaluation

Yes

Ability to generate high-quality patent text for two distinct tasks in machine-generated patent texts: claims-to-abstract generation, and the generation of the next claim given previous ones.

Subset

null

"Evaluating two distinct tasks in machine-generated patent texts: claims-to-abstract generation, and the generation of the next claim given previous ones"

Patent claims as input, paired with two outputs (machine-generated or human-written claims) that are evaluated through pairwise comparison by human experts.

null

Real task examples (e.g. GitHub issues)

400

Yes

patent domain, pairwise comparison, claim dependency

Specific criteria (items were taken from a larger set based on specified rules)

Free response (e.g. summary paragraph, executable code)

Human ratings (text quality, preference, NOT manual scoring of other metrics)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Win and Draw Rates

Yes

Error types, model, domain

null

https://github.com/ZoeYou/PatentEval

PatentEval

Contested

Yes

No

No comparisons made

No

Yes

(1) Expert annotation from patent lawyers; (2) Correlation analysis between human judgments and automated metrics; (3) Statistical evaluation of pairwise annotations; (4) Detailed error analysis

Win and draw rates from pairwise comparisons. For automated metrics and human judgment evaluation: Kendall's Tau

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

Real subtasks within the patent drafting process, though they don't cover the complete patent drafting workflow.

Composite phenomenon

Yes

null

No

Law

null

['Real task']

['Criterion']

['Free response']

['Human ratings']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Partial']

['Mean', 'Other']

yinNaturalLanguageCode2023

Natural Language to Code Generation in Interactive Data Science Notebooks

Include

null

A benchmark for data science tasks (natural language to code) in a Jupyter notebook. A benchmark with "realistic NL intents, rich notebook context, and a series of interrelated problems"

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Data science code generation

Yes

Clearly defines what a computational notebook is and that the aim of the phenomenon is to generate code for the next cell that satisfies the user's intent. Scope of intended code falls within "data wrangling" or "EDA"

Comprehensive

null

Generate code to fulfil the user intent for a specific cell, provided with the notebook history in-context. Code length is usually ~1-3 lines (pretty basic)

A natural language question specifying the user intent for the following cell.

All tasks involve pandas manipulations.

Real task examples (e.g. GitHub issues), Expert-crafted task examples (e.g. hand-written examples)

1,078

Yes

Data source (existing task vs newly created task)

Convenience sample (creators found a set of tasks that was readily accessible)

Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall), Execution-based evaluation (unit tests)

null

Industry

Yes

null

Test

null

Simple Mean

Yes

Task source (existing task vs new task)

pass@k (any correct answer in k trials)

https://github.com/google-research/arcade-nl2code/

ARCADE

Contested

Yes

No

The benchmark is itself realistic

No

Yes

- All tasks use pandas, which is realistic of real DS notebooks but still not 100% coverage of the phenomenon e.g. don't consider plotting problems.

Mean, error bars on figures in appendix.

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively), Partial real task (e.g. answering medical questions collected from real people)

null

Single cohesive phenomenon

Not applicable

null

Yes

Code Generation

null

['Real task', 'Expert-crafted']

['Convenience']

['Structured']

['Exact match', 'Reward']

['Contested']

['Yes']

['Realistic']

['Yes']

['Complete', 'Partial']

['Mean', 'Std']

zhangCABComprehensiveAttention2023

Quantifying Adaptability in Pre-trained Language Models with 500 Tasks

Include

null

We present a large-scale empirical study on LM adaptability using TASKBENCH500, a benchmark of 500 procedurally generated sequence modeling tasks. We evaluate three facets of adaptability, finding that: (1) adaptation methods vary in memorizing small datasets; (2) some show compositional adaptability to complex tasks; and (3) label distribution mismatches arise from differences in intrinsic label difficulty. Our results show that adaptability to new tasks can be systematically analyzed, similar to generalization.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

LM Adaptability

Yes

Adaptability: 'Adapting pre-trained language models (LMs) by finetuning their parameters or input prompts for downstream tasks' (Page 1)

Comprehensive

null

They define a set of atomic tasks, which are combined using a set of composition functions to produce more complex tasks. The atomic tasks include lexical tasks, factual tasks, and random relation tasks. The composition functions are word-level and sequential compositions.

For each task f, they construct a dataset D(f) = {(x_i, y_i)}, where x_i is sampled from the task's input distribution (e.g., most common words), and y_i is uniformly sampled from the set of valid outputs f(x_i), i.e., y_i ~ Unif(f(x_i)).

For evaluation, they measure the model’s average per-token accuracy on both training and test splits of the dataset D(f).

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

null

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically)

Short free response (e.g. single word or number), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Every task takes as input a word or word sequence, and outputs either a boolean value or a set of words/word sequences.

Simple Mean

No

null

https://github.com/facebookresearch/task_bench

TASKBENCH500

Contested

Yes

No

No comparisons made

No

Yes

The authors explicitly justify their synthetic task generation approach as a more controlled, interpretable, and generalizable way to study model adaptability, enabling them to identify which task attributes make learning easier or more difficult, something that is difficult to study with existing real-world datasets. 'For benchmarks built from collections of real-world datasets, the makeup and difficulty of these datasets is often difficult to characterize precisely: differences in annotation standards, annotation quality, and dataset size mean that models often exhibit very different performance on datasets designed to evaluate model performance on the same abstract task.' (Page 3)

null

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

No

null

No

NLP

Long Context

null

['Author-crafted']

['Convenience', 'Targeted']

['Short free response', 'Structured']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

null

kumarVisionlanguageModelsUnderstand2024

Do Vision-Language Models Understand Compound Nouns?

Include

null

We curate Compun, a novel benchmark with 400 unique and commonly used Compound Nouns (CN), to evaluate the effectiveness of VLMs in interpreting CNs. We perform an in-depth analysis to highlight CLIPs’ limited understanding of certain types of CNs. We present an alternative framework that moves beyond hand-written templates for text prompts widely used by CLIP-like models. We employ a Large Language Model to generate multiple diverse captions that include the CN as an object in the scene described by the caption. Our proposed method improves CN understanding of CLIP by 8.25% on Compun.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

VLMs' ability to understand compound nouns

Yes

A compound noun (CN) is a noun formed from two or more words combined to create a single noun with a new meaning. Interpreting the meaning of CNs by decoding the implicit semantic relation between their constituent nouns has attracted interest in NLP for decades. Though extensively studied in NLP, whether modern vision-language models (VLMs) understand CNs is under-explored. Their paper fills in this gap.

Comprehensive

null

Each instance in Compun corresponds to a unique compound noun and includes one image representing the compound noun (CN), along with two additional distractor images. These distractor images depict the individual constituent nouns that form the CN. Given the class name (or the CN), the task of a VLM is to retrieve (or select) the correct image among the distractors.

A unique compound noun, one image representing the compound noun (CN), two additional distractor images that depict the individual constituent nouns that form the CN.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

400

No

null

Random sample (creators defined a task space and sampled from it)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Multiple choice as in the VLM picks up one image out of three options

null

No

null

https://github.com/sonalkum/Compun

Compun

Widely-agreed

Yes

No

No comparisons made

No

null

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Single cohesive phenomenon

No

null

No

NLP

Understanding

null

['Author-crafted']

['Random']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Representative']

null

asaiBUFFETBenchmarkingLarge2024

BUFFET: Benchmarking Large Language Models for Few-shot Cross-lingual Transfer

Include

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Few-shot Cross-lingual Transfer

Yes

Few-shot cross-lingual transfer is defined as the ability to adapt models to a task in a new language using a limited number of training data in the target language.

Comprehensive

null

Converting all tasks into a unified text-to-text format, where models must generate appropriate outputs given inputs with k-shot examples in target languages.

Instruction, k-shot training and validation examples, test examples (input text and expected output)

null

Modified from another benchmark (e.g. translation into another language)

null

Yes

language, task type, data curation method, output format, languages per task

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF)

null

Mix (multiple authors from industry and academia)

Yes

null

hierarchical averaging

Yes

language, dataset, task type, resource level, fine-tuning/ICL

null

https://huggingface.co/datasets/BuffetFS/BUFFET

BUFFET

Widely-agreed

Yes

No

No comparisons made

No

null

Dataset score is calculated as a macro-average of the per-language score.

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

Yes

Language Modelling

In-context Learning

Multilinguality

['Another benchmark']

['Targeted', 'Criterion']

['Multiple choice', 'Short free response', 'Free response']

['Exact match', 'Soft match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Representative']

['Mean']

hareshClevrSkillsCompositionalLanguage2024

ClevrSkills: Compositional Language And Visual Reasoning in Robotics

Include

null

We ask the question: if the models are taught the low-level capabilities, can they compose them in novel ways to achieve high-level tasks like cleaning the table without having to be explicitly taught so? To this end, we present ClevrSkills - a benchmark suite for compositional reasoning in robotics. The dataset contains trajectories generated on a range of robotics tasks with language and visual annotations as well as multi-modal prompts as task specification. We benchmark multiple different VLM baselines on ClevrSkills and show that even after being pre-trained on many tasks, these models fail on compositional reasoning in robotics tasks.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Compositional reasoning/generalization in robotics

Yes

Compositional generalization is a hallmark feature of human intelligence. Unlike any other animals, humans can receive instructions in natural language and successfully perform previously unseen tasks with minimal to no task-specific learning or adaptation.' (Page 1)

null

In ClevrSkills, they benchmark robotic models on a set of simple manipulation tasks, such as pick, place, throw, touch and push, and evaluate their ability to generalize to complex tasks based on these low-level capabilities. Tasks are organized into three levels (L0 → L1 → L2), where higher-level tasks build on skills from lower levels to assess compositional reasoning.

A task prompt (plain text or multi-modal), a sequence of action labels (skill traces and language annotations), corresponding RGB observations from multiple camera views, key visual steps, and dense rewards over time. It represents one complete trajectory for solving a specific manipulation task.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

330k trajectories on 33 tasks

Yes

For each of 330k trajectory, it contains many types of annotation, including language, action classes, bounding boxes for objects, visibility annotations, key-steps, rewards (for offline RL), camera parameters and more.

Targeted items (creators defined a task space and chose tasks within it strategically)

Extended interaction (e.g. conversation, calling an API and processing the response)

Exact Match (accuracy, F1, precision, recall)

Per-task success rates

null

Industry

Yes

null

Test

null

A series of robot actions

Simple Mean

Yes

Success rates on L0, L1 and L2 tasks

null

https://github.com/Qualcomm-AI-research/ClevrSkills

ClevrSkills

Contested

Yes

No

No comparisons made

No

null

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

No

Reasoning

Compositional

null

['Author-crafted']

['Targeted']

['Interaction']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['No']

['Partial']

null

hanMedSafetyBenchEvaluatingImproving2024

MedSafetyBench: Evaluating and Improving the Medical Safety of Large Language Models

Include

null

The paper introduces the first benchmark dataset designed to measure the medical safety of LLMs. It uses the dataset to evaluate and improve the medical safety of LLMs using fine-tuning.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

medical safety

Yes

"we define an LLM to be aligned with medical safety standards if its output is not only accurate but also consistent with the AMA (American Medical Association)'s Principles of Medical Ethics." (page 3)

Comprehensive

null

They prompt LLMs with harmful medical requests and evaluate the harmfulness of their responses

harmful request, category of harm

null

LLM-generated task examples (e.g. Filtered from responses to a prompt)

900

Yes

principle violated

Random sample (creators defined a task space and sampled from it)

Free response (e.g. summary paragraph, executable code)

LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics)

null

They validate the LLM-generated dataset by conducting a user study with 25 doctors to check that the generated prompts violate one of the nine principles of medical ethics.

Academia

Yes

null

Test, Train

Train: 900

null

Simple Mean

Yes

By LLM used to generate the prompts

null

https://github.com/AI4LIFE-GROUP/med-safety-bench

MedSafetyBench

Contested

Yes

No

No comparisons made

No

null

They indirectly address it by noting in their discussion: "In practice, one could consider introducing nuance to the definition. For example, levels of acceptable risk may vary among medical subspecialties (e.g., emergency medicine vs. neurological surgery vs. dermatology) and based on a patient’s condition and personal preference (e.g., a patient with a condition that has no established treatment options may be more willing to try risky experimental procedures). Aligning LLMs to account for different levels of acceptable risk and be tailored to different medical subspecialties is a future research direction"

simple mean, standard error of the mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

No

Medicine

null

['LLM-generated']

['Random']

['Free response']

['LLM-as-a-Judge']

['Contested']

['Yes']

['No comparison made']

['']

['Partial']

['Mean']

shivagundeLargerProbesTell2023

Larger Probes Tell a Different Story: Extending Psycholinguistic Datasets Via In-Context Learning

Include

null

Psycholinguistic datasets for negation and role reversal, which extend existing smaller benchmarks using GPT-3. Evaluation of multiple LMs on these extended benchmarks reveals performance drops.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

negation, role reversal

Yes

Negation sensitivity (LMs ability to understand negation); Role reversal (ability to understand reversing semantic roles)

Subset

null

Accurately predicting masked/target words in constructed sentence pairs - given negation and role reversal.

null

Pair of sentences, e.g., an affirmative sentence and its negated counterpart.

Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

3000

No

null

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), Distribution (perplexity, calibration, correlation)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

negation, role reversal, models, dataset versions

pass@k (any correct answer in k trials)

https://github.com/text-machine-lab/extending_psycholinguistic_dataset

NEG-1500-SIMP, ROLE-1500

Widely-agreed

Yes

No

Yes

"We validated our new datasets through human evaluation. For this, we randomly selected 100 samples from each of the extended datasets"

Simple means, McNemar test, Minimum detectable effect (MDE)

Model access required (e.g. logits)

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

No

NLP

Understanding

null

['Another benchmark', 'LLM-generated']

['Targeted', 'Criterion']

['Short free response']

['Exact match', 'Distribution']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Mean', 'Other']

wuSmartPlayBenchmarkLLMs2024

SMARTPLAY : A BENCHMARK FOR LLMS AS INTELLIGENT AGENTS

Include

null

This paper introduces SmartPlay, a benchmark for assessing LLMs as intelliget agents using 6 different games including Rock-Paper-Scissors, Tower of Hanoi, Minecraft. Each game features a unique setting, providing up to 20 evaluation settings and infinite environment variations. Each game in SmartPlay uniquely challenges a subset of 9 important capabilities of an intelligent LLM agent, including reasoning with object dependencies, planning ahead, spatial reasoning, learning from history, and understanding randomness.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

LLMs as intelligent agents

No

While the authors do not explicitly provide a definition for intelligent agents, they provide some key properties of agents from which the target phenomenon aims to be captured—i.e., LLM agents as systems capable of long-horizon planning, probabilistic reasoning, spatial reasoning to understand the 3D world, and learning from interactions or mistakes. This is further decomposed into 9 measurable abilities; long text understanding, reasoning, instruction following, planning, generalization, understanding the odds, learning from interactions, error/mistake handling and spatial reasoning.

Comprehensive

null

An LLM is provided with environment-specific inputs; either textual descriptions or visual descriptions (via natural language)—along with manuals containing background knowledge, rules, and examples. The LLM must then interact with the environment by selecting actions from a predefined action space to achieve task objectives across multiple trials or rollouts.

A task item is a description of a game, the rules, actions, environment and expected behaviour.

A task item consists of a description of the game, its rules, available actions, environment state, and the expected agent behavior.

Real task examples (e.g. GitHub issues)

6 games with 20 different evaluation settings

Yes

Each task item is annotated with the abilities required i.e. long text understanding, reasoning, instruction following, planning, generalization, understanding the odds, learning from interactions, error/mistake handling and spatial reasoning.

Convenience sample (creators found a set of tasks that was readily accessible)

Extended interaction (e.g. conversation, calling an API and processing the response)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

Across each game in the task

null

https://github.com/microsoft/SmartPlay

SMARTPLAY

Widely-agreed

Yes

No

No comparisons made

The benchmark is itself realistic

Yes

Somewhat

The authors use real-world-inspired games to evaluate LLM capabilities as intelligent agents, but the extent to which success in these environments generalizes to real-world agentic behavior remains an open question.

Scores are normalised relative to human performance.

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

null

Composite phenomenon

Yes

null

No

Agents

null

['Real task']

['Convenience']

['Interaction']

['Exact match']

['Widely-agreed']

['Yes']

['Realistic']

['Yes']

['Complete']

['Other']

liWMDPBenchmarkMeasuring2024

The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning

Include

null

The paper introduces a benchmark with questions that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security. It also introduces a state-of-the-art unlearning method to reduce model performance on the benchmark.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

hazardous capabilities related to biosecurity, cybersecurity, and chemistry knowledge

No

Knowledge of biosecurity, cybersecurity, and chemistry that could be hazardous

Subset

null

MCQ questions with "knowledge that is a precursor, neighbor, or component of the hazardous knowledge we wish to remove" (page 4)

MCQ question, options, correct answer

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Expert-crafted task examples (e.g. hand-written examples)

3,668

Yes

topic area (biology, cyber, chemistry)

Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Simple Mean

Yes

topic area (bio, cyber)

null

https://huggingface.co/datasets/cais/wmdp

The WMDP Benchmark

Contested

No

Yes

No

No comparisons made

No

null

simple mean

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

No

Alignment

Safety

null

['Author-crafted', 'Expert-crafted']

['Targeted']

['Multiple choice']

['Exact match']

['Contested']

['No']

['Yes']

['No comparison made']

['No']

['Representative']

['Mean']

renValueBenchComprehensivelyEvaluating2024

ValueBench: Towards Comprehensively Evaluating Value Orientations and Understanding of Large Language Models

Include

null

This paper introduces ValueBench, a benchmark for evaluating value orientations and value understanding in LLMs, grounded in realistic human-AI interactions to test for value orientations, along with new tasks for evaluating value understanding in an open-ended value space.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Value orientation and value understanding

Yes

Values are concepts or beliefs about desirable end states or behaviors that transcend specific situations. Value Orientation is the extent to which an LLM exhibits preferences or inclinations toward specific human values — i.e., how aligned its responses are with particular value statements or stances. Value Understanding is the extent to which an LLM can recognise, interpret, and reason about human values — including identifying relationships between values, inferring values from behaviours or statements, and generating expressions that reflect particular values.

Comprehensive

null

In value orientation, an LLM is given a value loaded question converted from psychometric statements abd asked to respond with advice. In value understanding, an LLM is prompted to identify relevant values on both positive and negative samples. For each value pair, the LLMs are required to sequentially output the definition of both values, a brief explanation of their relationship, the corresponding relationship label, and a final assessment of relevance (1 if relevant and 0 otherwise). An LLM can also be presented with a value name and is requireed to generate a value-reflecting statement or a pairs of values and required to identify semantic relationships.

For value orientation a single task item is a rephrased value statement to advice-seeking question. For value understanding, the task items can be a behavioural statement, a value name or pairs of values.

null

Real task examples (e.g. GitHub issues)

44 psychometric inventories, 453 value dimensions and 1989 value orientation questions

No

null

Convenience sample (creators found a set of tasks that was readily accessible)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), LLM post-processing (extracting answers, reformatting for automated scoring)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

Across each value orientation and subtask dataset

null

https://github.com/Value4AI/ValueBench

ValueBench

Widely-agreed

Yes

No

Yes

No

No comparisons made

No

Yes

The benchmark contains questions sourced from psychometric analysis tests and shows correlations in the results for tests that examine similar behaviours.

null

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

No

Alignment

null

['Real task']

['Convenience']

['Short free response']

['Exact match', 'LLM post-processing']

['Widely-agreed']

['Yes']

['No']

['No comparison made']

['Yes']

['Partial']

null

wangDecodingTrustComprehensiveAssessment2023

DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models

Include

null

The paper introduces a comprehensive trustworthiness evaluation for large language models with a focus on GPT-4 and GPT-3.5. The benchmark introduced considers diverse perspectives including toxicity, stereotype bias, adversarial robustness, out-of-distribution robustness, robustness on adversarial demonstrations, privacy, machine ethics, and fairness.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Trustworthiness

No

"toxicity, stereotype bias, adversarial robustness, out-of-distribution robustness, robustness on adversarial demonstrations, privacy, machine ethics, and fairness." (page 2)

Comprehensive

null

Presenting LLMs with different scenarios that invoke one of the subsets of trustworthiness (e.g., stereotypical input) and assessing output.

Varies by task (the benchmark includes many tasks and datasets) but seems to include user prompt, system prompt, choices (when prompts are MCQ)

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

null

Yes

system prompt, prompt template

Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), Distribution (perplexity, calibration, correlation)

null

Mix (multiple authors from industry and academia)

Yes

null

Simple Mean, Weighted Mean

Yes

different system prompts, normal vs challenging prompts, different adversarial demonstrations, different adversarial text generation strategies, different demographic groups and stereotype topics, different types of PII, different sensitive attributes

pass@k (any correct answer in k trials)

https://github.com/AI-secure/DecodingTrust

DecodingTrust

Contested

Yes

No

indirectly address it

In the limitations section: "Subjectivity. Trustworthiness perspectives such as toxicity, stereotype bias, machine ethics, and fairness involve subjectivity and should be human-centric in their definitions and evaluations. Here we aim to provide our objective observations, and leave the analysis of how these metrics are aligned with human as an important future work to explore model behaviors based on human understanding"

simple mean, weighted mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people), Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

They provide the total # of prompts used for each trustworthiness subset (in the context of calculating computational costs) but since they run a lot of variations (e.g., different system prompts), it's unclear what the size of the bare benchmark is. This is in Appendix K.

Yes

Alignment

null

['Author-crafted', 'Another benchmark', 'LLM-generated']

['Targeted']

['Multiple choice', 'Short free response', 'Free response']

['Exact match', 'Distribution']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Partial', 'Representative']

['Mean']

chakrabortyCounterTuringTest2023

Counter Turing Test (CT2): AI-Generated Text Detection is Not as Easy as You May Think – Introducing AI Detectability Index

Include

null

Counter Turing Test (CT2) is a benchmark to evaluate the robustness of AI-generated text detection techniques. The AI Detectability Index (ADI) is a metric to rank LLMs according to how detectable their outputs are as machine-generated versus human-written.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

AI-generated text detection

Yes

"If a work’s traditional elements of authorship were produced by a machine, the work lacks human authorship", Detectability of AI-generated text

Comprehensive

null

The Counter Turing Test (CT2) evaluates the robustness of AI-generated text detection techniques: (i) watermarking, (ii) perplexity estimation, (iii) burstiness estimation, (iv) negative log-likelihood curvature, (v) stylometric variation, and (vi) classifier-based approaches.

Pair of texts on the same topic (one human-written and one AI-generated).

null

Real task examples (e.g. GitHub issues), Procedurally-generated task examples (e.g. Creating instances from a template)

100000

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall), Distribution (perplexity, calibration, correlation)

null

Mix (multiple authors from industry and academia)

No, no link is provided

null

Test

null

Formula combining perplexity and burstiness

Yes

Detection method, LLM used

null

CT2

Widely-agreed

Yes

No

Yes

No

null

mean, standard deviation, entropy calculations, z-scores, p-values, bootstrapping, Le Cam's lemma, multiplicative damping factors

Model access required (e.g. logits)

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

No

NLP

Detection

null

['Real task', 'Procedurally-generated']

['Convenience', 'Criterion']

['Multiple choice']

['Exact match', 'Distribution']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Representative']

['Mean', 'Std', 'Tests', 'Other']

tuWaterBenchHolisticEvaluation2024

WaterBench: Towards Holistic Evaluation of Watermarks for Large Language Models

Include

null

WaterBench is a benchmark for evaluating LLM watermarks across detection and generation quality. The paper also presents a hyper-parameter search method to control watermarking strength, and automatic evaluation using GPT4-Judge. The dataset is publicly available and human-validated.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

LLM watermarks

Yes

A watermarked LLM generates texts "with a biased distribution of tokens, which distinguishes it from unwatermarked texts, ... , the goal of watermarking is to achieve high detection accuracy while maintaining the generation quality" (1517).

Comprehensive

null

WaterBench consists of 9 tasks with 5 unique task settings, spanning "a wide range of input and output length" (1520). The first setting is Short Input, Short Answer, and has two tasks to evaluate factual knowledge. The second setting is Short Input, Long Answer, with two Long-form QA tasks. The third category is Long Input, Short Answer, with reasoning and code completion tasks. The fourth setting is Long Input, Long Answer, with two summarization tasks. The last setting is open-ended generation, where the task is instruction-following.

Each task is sourced from a different dataset and has its own features.

null

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language)

2405

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall), Human ratings (text quality, preference, NOT manual scoring of other metrics), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics), Generation Metric and Generation Quality Drop are never explicitly defined in the paper.

The paper defines watermarking strenth as the true positive rate to ensure all watermarks are of similar intensity during evaluation. WaterBench also uses GPT4-Judge, which "measures which model's output the GPT-4 system prefers when shown two responses for the same instruction" (1521). The paper reports the "True Positive Rate, True Negative Rate, Generation Metric, and Generation Quality Drop for all tasks" (1523). 100 responses are randomly sampled for human annotation as well.

Each task is sourced from a different dataset.

Academia

Yes

null

Test

null

Simple Mean

No

null

https://github.com/THU-KEG/WaterBench

WaterBench

Widely-agreed

Yes

No

No comparisons made

The benchmark is itself realistic

No

Yes

The authors highlight that evaluating watermark must "evaluate the generation and detection" methods, and ensure fair comparisons between watermarking methods (1517). Additionally, the authors highlight that the tasks must be diverse, go beyond "text completion," and measure "generation quality" and alignment with "human preferences" (1517-1518).

True Positive Rate, True Negative Rate, Generation Metric and Generation Quality Drop

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

null

Single cohesive phenomenon

No

null

No

NLP

Detection

null

['Real task', 'Another benchmark']

['Convenience', 'Targeted', 'Criterion']

['Short free response', 'Free response']

['Exact match', 'Human ratings', 'LLM-as-a-Judge', 'LLM-as-a-Judge']

['Widely-agreed']

['Yes']

['Realistic']

['Yes']

['Complete']

['Mean']

luoMMMRSMultimodalMultiGSD2024

MMM-RS: A Multi-modal, Multi-GSD, Multi-scene Remote Sensing Dataset and Benchmark for Text-to-Image Generation

Include

null

MMM-RS is a large, multi-modal multi-GSD, and multi-scene remote sensing text-to-image generation benchmark. It is publicly available, aggregated and filtered from existing datasets, and contains information-rich captions.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

remote sensing text-to-image generation

No

Remote sensing image generation is the ability to prompt a multimodal model to generate a high-quality remote sensing images.

Comprehensive

null

A multimodal model is given an information-rich text prompt and must generate the described remote sensing image.

An remote sensing image and an information-rich text prompt, specific to its image modality. For example, the prompt may contain satellite type, weather type, category, resolution, subject, etc.

null

Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

2,103,273

No

=

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Image

Distribution (perplexity, calibration, correlation)

The paper uses Frechet Inception Distance (FID) and Inception Score (IS).

"The MMM-RS dataset is derived from 9 publicly available RS datasets: MRSSC2.0 [16], Inria [19], NaSC-TG2 [45], GID [28], WHU-OPT-SAR [14], HRSC2016 [40], TGRS-HRRSD [42], fMoW [5], and SEN1-2 [25]" (4). It contains images across three modalities: RGB, Synthetic Aperture Radar, and Near Infrared. Multi-scene remote sensing images are synthesized at different scales and weather conditions using physics models and multimodal models. The process is outlined in Figure 4 on Page 6.

Academia

Yes

null

Test, Train

Optional train set is defined at 200,000 samples.

null

Simple Mean

No

null

https://github.com/ljl5261/MMM-RS

MMM-RS (Multi-modal, Multi-GSD, Multi-scene Remote Sensing)

Widely-agreed

Yes

No

No comparisons made

The benchmark is itself realistic

No

Yes

The authors highlight that a remote sensing text-to-image generation dataset should be multimodal across data and image types, and be information-rich.

Simple mean

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

null

Single cohesive phenomenon

Not applicable

null

No

Grounding

null

['Another benchmark', 'LLM-generated']

['Targeted', 'Criterion']

['Free response']

['Distribution']

['Widely-agreed']

['Yes']

['Realistic']

['Yes']

['Complete']

['Mean']

devriesDUMBBenchmarkSmart2023

DUMB: A Benchmark for Smart Evaluation of Dutch Models

Include

null

DUMB, a benchmark for evaluating Dutch language models across nine tasks. The authors propose Relative Error Reduction (RER) for better cross-task comparison and evaluate pre-trained models, finding that current Dutch models underperform while identifying strategies for future model improvements.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Dutch language understanding

No

null

Subset

null

Word-level tasks (POS tagging, NER), word-pair tasks (word sense disambiguation, pronoun resolution), sentence-pair tasks (causal reasoning, natural language inference), and document-level tasks (sentiment analysis, abusive language detection, question answering).

Task-specific, e.g., input text (word, sentence, or document) with corresponding label/target output.

null

Modified from another benchmark (e.g. translation into another language)

train (sum): 283,112

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number), Free response (e.g. summary paragraph, executable code)

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test, Train, Validation

test (sum): 23,406 ; dev (sum): 17,152

null

Relative Error Reduction (RER)

Yes

Scores for each of the nine individual tasks; grouped average scores by model type, model size, pre-training language

null

https://github.com/wietsedv/dumb

DUMB

Contested

Yes

No

null

Binomial mixed effects regression models

Outputs alone

Representative task (e.g. answering medical licensing exam questions), Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

Standard NLP evaluation benchmarks rather than complete real-world applications, but they are important for real-world Dutch language processing applications.

Composite phenomenon

Yes

null

No

NLP

Understanding

Multilinguality

['Another benchmark']

['Convenience', 'Criterion']

['Short free response', 'Free response']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['No']

['Representative', 'Constructed']

['Tests']

hsiehSugarCrepeFixingHackable2023

SugarCrepe: Fixing Hackable Benchmarks for Vision-Language Compositionality

Include

null

We introduce SUGARCREPE, a new benchmark for vision-language compositionality evaluation. We employ large language models, instead of rule-based templates used in previous benchmarks, to generate fluent and sensical hard negatives, and utilize an adversarial refinement mechanism to maximally reduce biases.' (abstract)

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Compositional understanding of vision-language models

Yes

Through compositional reasoning, humans can comprehend new scenes and describe those scenes by composing known atoms. For instance, compositionality allows people to differentiate between a photo of “a girl in white facing a man in black” and “a girl in black facing a man in white”. Vision-language research has sought to develop models that can similarly comprehend scenes and express them through compositional language.' (page 1)

Comprehensive

null

This is formulated as an image-to-text retrieval task. It evaluates a vision-language model’s ability to distinguish the correct caption for an image from a closely matched, compositionally altered hard negative, thereby testing its compositional understanding of visual scenes.

An image paired with two captions: one positive caption that correctly describes the image, one hard negative caption that is compositionally similar but incorrect.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

7512

Yes

It provides types of compositional perturbations applied to generate hard negatives. These include REPLACE-OBJ, REPLACE-ATT, REPLACE-REL, SWAP-OBJ, SWAP-ATT, ADD-OBJ, and ADD-ATT, which specify the type of atomic concept (object, attribute, relation) and the operation used (replace, swap, add).

Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Multiple choice as each image presents two captions (a positive and a hard negative), and the model choose which one correctly describes the image.

null

Yes

Seven fine-grained hard negative types under three categories of REPLACE, SWAP, and ADD: REPLACE: REPLACE-OBJ (object substitutions), REPLACE-ATT (attribute substitutions); REPLACE-REL (relation substitutions); SWAP: SWAP-OBJ (object swaps), SWAP-ATT (attribute swaps); ADD: ADD-OBJ (adding an object), ADD-ATT (adding an attribute).

null

https://github.com/RAIVNLab/sugar-crepe

SugarCrepe

Contested

Yes

No

Yes

The benchmark was proposed to address the construct validity failures in prior benchmarks (e.g., ARO, CREPE) by showing that a blind model can even succeed without using visual input. They address it in their benchmark by improving hard negative generation using ChatGPT, followed by adversarial refinement, to ensure captions differed only in compositional meaning and remove any biases.

null

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Single cohesive phenomenon

Not applicable

null

No

Reasoning

Compositional

null

['Author-crafted']

['Targeted']

['Multiple choice']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Representative']

null