NoLiMa: Long-Context Evaluation Beyond Literal Matching

This repository contains the data associated with our ICML 2025 paper, "NoLiMa: Long-Context Evaluation Beyond Literal Matching".

Abstract

Recent large language models (LLMs) support long contexts ranging from 128K to 1M tokens. A popular method for evaluating these capabilities is the needle-in-a-haystack (NIAH) test, which involves retrieving a "needle" (relevant information) from a "haystack" (long irrelevant context). Extensions of this approach include increasing distractors, fact chaining, and in-context reasoning. However, in these benchmarks, models can exploit existing literal matches between the needle and haystack to simplify the task. To address this, we introduce NoLiMa, a benchmark extending NIAH with a carefully designed needle set, where questions and needles have minimal lexical overlap, requiring models to infer latent associations to locate the needle within the haystack. We evaluate 12 popular LLMs that claim to support contexts of at least 128K tokens. While they perform well in short contexts (<1K), performance degrades significantly as context length increases. At 32K, for instance, 10 models drop below 50% of their strong short-length baselines. Even GPT-4o, one of the top-performing exceptions, experiences a reduction from an almost-perfect baseline of 99.3% to 69.7%. Our analysis suggests these declines stem from the increased difficulty the attention mechanism faces in longer contexts when literal matches are absent, making it harder to retrieve relevant information.

Results

Models	Claimed Length	Effective Length	Base Score (×0.85: Thr.)	1K	2K	4K	8K	16K	32K
GPT-4o	128K	8K	99.3 (84.4)	98.1	98.0	95.7	89.2	81.6	69.7
Llama 3.3 70B	128K	2K	97.3 (82.7)	94.2	87.4	81.5	72.1	59.5	42.7
Llama 3.1 405B	128K	2K	94.7 (80.5)	89.0	85.0	74.5	60.1	48.4	38.0
Llama 3.1 70B	128K	2K	94.5 (80.3)	91.0	81.8	71.2	62.7	51.8	43.2
Gemini 1.5 Pro	2M	2K	92.6 (78.7)	86.4	82.7	75.4	63.9	55.5	48.2
Jamba 1.5 Mini	256K	<1K	92.4 (78.6)	76.3	74.1	70.8	62.2	52.7	43.6
Command R+	128K	<1K	90.9 (77.3)	77.0	73.5	66.3	39.5	21.3	7.4
Llama 4 Maverick 🆕	1M	2K	90.1 (76.6)	81.6	78.3	68.8	⏳	⏳	⏳
Gemini Flash 2.0 🆕	1M	4K	89.4 (76.0)	87.7	87.5	77.9	64.7	48.2	41.0
Gemma 3 27B 🆕	128K	<1K	88.6 (75.3)	73.3	65.6	48.1	32.7	20.2	9.5
Mistral Large 2	128K	2K	87.9 (74.7)	86.1	85.5	73.3	51.5	32.6	18.7
Claude 3.5 Sonnet	200K	4K	87.6 (74.4)	85.4	84.0	77.6	61.7	45.7	29.8
Gemma 3 12B 🆕	128K	1K	87.4 (74.3)	74.7	61.8	39.9	27.4	16.8	7.3
Gemini 1.5 Flash	1M	<1K	84.7 (72.0)	68.6	61.6	51.0	44.4	35.5	28.6
GPT-4o mini	128K	<1K	84.9 (72.2)	67.7	58.2	44.1	32.6	20.6	13.7
Llama 4 Scout 🆕	10M	1K	81.7 (69.4)	72.3	61.8	50.8	35.5	26.9	21.6
Llama 3.1 8B	128K	1K	76.7 (65.2)	65.7	54.4	44.1	31.9	22.6	14.2
Gemma 3 4B 🆕	128K	<1K	73.6 (62.6)	50.3	35.3	16.4	7.5	2.3	0.9

This table presents the performance results of selected models on NOLIMA tests. The base score represents a model’s accuracy on the task at short contexts (250, 500, and 1K) and serves as a controlled reference to measure performance degradation at longer contexts. The effective length is defined as the longest context where a model maintains at least 85% of its base score. Scores above this threshold are underlined, while scores dropping below 50% of the base score are italicized.

✨ Updates:

• [2025-04-10]: Added evaluation results on Gemma 3 models (4B/12B/27B), Gemini 2.0 Flash, and Llama 4 Scout. (Llama 4.0 Maverick evaluation in progress... ⏳)

NoLiMa-Hard Results

Models	Base Score	4K	8K	16K	32K
Llama 3.3 70B
- w/o CoT	98.3	55.5	37.2	16.7	8.9
- w/ CoT	97.1	73.0	51.2	31.8	10.1
Reasoning Models
GPT-o1	99.9	92.0	78.0	60.1	31.1
GPT-o3 Mini	98.8	52.8	36.9	25.5	18.9
DeepSeek R1-Distill-Llama-70B	99.9	91.4	75.5	49.4	20.7

This table presents the performance results of selected reasoning models on NoLiMa-Hard, a subset of the original NoLiMa needle set containing the 10 most challenging question-needle pairs from previous evaluations. Scores dropping below 50% of the base score are in italic.

Evaluation

This HuggingFace repository contains all the necessary data--including the NoLiMa needle set and the haystacks--to evaluate models on the NoLiMa benchmark.

To access the evaluation script and more information, please refer to the NoLiMa GitHub repository.

Dataset Structure

The dataset is structured as follows:

• haystack/ : Contains the haystack data
  • books.tar.gz : Contains the books used to generate the haystacks. It can also be used to create new shuffled haystacks.
  • rand_shuffle/ : Contains the shuffled haystacks that were used in the evaluation.
• needlesets/ : Contains the NoLiMa needle sets:
  • needle_set.json : The main NoLiMa needle set.
  • needle_set_hard.json : The NoLiMa-Hard needle set; a subset of the main needle set containing the 10 most challenging question-needle pairs.
  • needle_set_ONLYDirect.json : The main needle set with only direct questions.
  • needle_set_MC.json : The main needle set formatted as multiple-choice questions.
  • needle_set_w_CoT.json : The main needle set with CoT task templates.
  • needle_set_w_distractor.json : The main needle set with distractors.

GitHub Repo & Paper

For more information about NoLiMa, refer to:

• 📄 Paper: "NoLiMa: Long-Context Evaluation Beyond Literal Matching"
• 🔗 GitHub Repo: NoLiMa GitHub repository

License

The evaluation code and needle set data is licensed under the Adobe Research License. The license prohibits commercial use and allows non-commercial research use. For details about the haystack data, please refer to the haystack/LICENSES.md file.

Cite

If you use the NoLiMa dataset, filtering pipeline, or code, please cite the paper:

@inproceedings{modarressi2025nolima,
  title={NoLiMa: Long-Context Evaluation Beyond Literal Matching},
  author={Modarressi, Ali and Deilamsalehy, Hanieh and Dernoncourt, Franck and Bui, Trung and Rossi, Ryan A. and Yoon, Seunghyun and Schütze, Hinrich},
  booktitle={Forty-second International Conference on Machine Learning},
  year={2025},
  url={https://arxiv.org/abs/2502.05167}
}