Knowledge or Reasoning

Analysis

Why This Paper Matters

This paper addresses a critical gap in evaluating reasoning-enhanced LLMs: the over-reliance on final-answer accuracy. By decomposing reasoning into knowledge correctness and reasoning quality, it offers a more nuanced view of model behavior, especially in high-stakes domains like medicine where both factual accuracy and logical coherence are vital. The finding that SFT improves accuracy at the cost of reasoning quality (38.9% drop in InfoGain) is a stark warning for practitioners who prioritize benchmark scores over process transparency.

Moreover, the paper's demonstration that RL can prune inaccurate knowledge in medical reasoning suggests a path toward more reliable AI systems. This is particularly relevant as models like OpenAI-o1 and DeepSeek-R1 become more prevalent, yet their internal reasoning remains opaque. The work underscores the need for domain-specific training strategies rather than assuming general reasoning abilities transfer universally.

Technical Contributions

• Fine-grained evaluation framework: Introduces Knowledge Index (KI) to measure factual correctness and Information Gain (InfoGain) to quantify reasoning quality, moving beyond final-answer accuracy.
• Decomposition of thinking trajectories: Explicitly separates knowledge and reasoning components in step-by-step outputs, enabling targeted analysis.
• Comparative study of training paradigms: Evaluates SFT and RL on R1-distilled and base Qwen models across medical and math domains, revealing distinct effects.
• Key insight on RL: Shows RL enhances reasoning by eliminating irrelevant or incorrect knowledge, rather than just optimizing for answer correctness.

Results

• SFT improves final-answer accuracy in both medical and math domains but reduces InfoGain by an average of 38.9% compared to untrained models.
• General reasoning abilities from R1-distilled models do not transfer effectively to the medical domain via either SFT or RL.
• In the medical domain, SFT remains necessary for domain knowledge, but RL is crucial for improving reasoning quality and knowledge correctness.
• RL prunes inaccurate or irrelevant knowledge from reasoning paths, leading to better overall performance.

Significance

This research has broad implications for the AI field, particularly for deploying LLMs in sensitive applications like healthcare. It challenges the prevailing focus on final-answer metrics and advocates for evaluating reasoning processes. The finding that RL can enhance reasoning quality without sacrificing accuracy offers a practical guideline for model training. Future work may extend this framework to other domains and explore hybrid approaches combining SFT and RL to balance knowledge acquisition and reasoning fidelity.