📜 Paper Overview

We introduce ORION, a framework designed to teach language models to reason efficiently using a compact, symbolic language-of-thought representation. Instead of generating long, often redundant chain-of-thought explanations, we train models to express reasoning through concise, structured symbolic steps. Our approach combines supervised learning with a novel reinforcement learning objective to optimize both correctness and efficiency.

⏳ Evolutionary Context

Reasoning in large language models has evolved from prompt-based chain-of-thought, to supervised step-by-step reasoning, to reinforcement-learning-driven verifiable reasoning. We position ORION as the next step: efficient symbolic reasoning, moving toward future systems capable of scalable, faithful, and cost-effective problem solving with minimal redundancy.

🧬 Core Framework

Symbolic Reasoning Format

We formalize a Mentalese-style representation that encodes reasoning using atomic operations. This removes natural-language verbosity and keeps every operation semantically necessary.

Training Pipeline

• Foundation Stage: Supervised fine-tuning on symbolic reasoning traces from our MentaleseR dataset.
• RL Stage: Reinforcement learning with verifier feedback using Shorter Length Preference Optimization (SLPO), which encourages concise yet correct reasoning without rigid length constraints.

Evaluation System

We benchmark ORION on high-difficulty math reasoning tasks such as AIME, AMC, and MATH-500, measuring trace length, accuracy, inference cost, and faithfulness.

Capability Optimization

Our RL objective adaptively balances correctness and compactness, enabling models to keep reasoning steps minimal while preserving validity.

Efficiency Engineering

We implement a full pipeline optimized for deployment: shorter traces yield 4–16× fewer tokens, 5× lower latency, and 7–9× lower compute cost.

Interpretability & Faithfulness

Structured symbolic traces allow transparent inspection of each reasoning step, improving debuggability and model reliability.

🔭 Key Insights

• Most chain-of-thought traces are highly redundant; symbolic reasoning eliminates unnecessary natural-language overhead.
• Conciseness improves both model efficiency and faithfulness, reducing hallucinated or irrelevant steps.
• Reward shaping through SLPO is more flexible and robust than rigid length constraints.
• Efficient reasoning is essential for real-time agents and resource-constrained deployments.

💡 Value Proposition

ORION provides a clear technical pathway for efficient reasoning in LLMs. By combining symbolic thinking with reinforcement learning, we demonstrate that models can achieve strong accuracy with far fewer reasoning tokens. This work serves as both a conceptual shift away from verbose chain-of-thought and a practical system enabling faster, cheaper, and more interpretable reasoning, suitable for both academic research and real-world applications.