Evaluation of Reasoning Model Faithfulness: A Benchmark for Identifying 'Correct Answer, Incorrect Reasoning'

Introduces an open-source benchmark called reasoning-faithfulness-eval maintained by avilog, which aims to evaluate the chain-of-thought faithfulness of reasoning models. Through three scenarios—clean prompts, suggestive clues, and misleading clues—it detects whether models arrive at answers based on genuine correct reasoning, addressing the reasoning hallucination problem of 'correct answer, incorrect reasoning'. The project source is GitHub, released on June 5, 2026.

With the rise of reasoning models like OpenAI o1 and DeepSeek R1, chain-of-thought capabilities have improved interpretability, but new issues have emerged: models may get correct answers through guessing or pattern matching, yet their displayed reasoning process is wrong or fabricated ('correct answer, incorrect reasoning'). This type of reasoning hallucination is harder to detect.

The reasoning-faithfulness-eval benchmark designs three comparative scenarios: 1. Clean prompt scenario (standard question with no extra clues); 2. Suggestive clue scenario (embedded with correct prompts); 3. Misleading clue scenario (added with incorrect information). By comparing performance across scenarios, it judges whether the model's reasoning is based on internal logic rather than superficial clues.

The benchmark focuses on core dimensions: 1. Matching degree between answer accuracy and reasoning accuracy; 2. Clue sensitivity (utilizing valid clues and resisting misleading ones); 3. Consistency of reasoning process (detecting contradictions or errors in intermediate steps).

Implications of this benchmark for development: 1. The faithfulness of the reasoning process is as important as answer correctness; 2. Models are prone to being misled, so robustness needs to be enhanced (e.g., adversarial sample training); 3. Model comparisons should consider performance under misleading information.

Application developers can use this framework to understand model reasoning characteristics; if a model has poor resistance to misleading information, prompt engineering protection needs to be strengthened. The evaluation method can be extended to fields like code generation, scientific Q&A, and medical diagnosis to identify fabricated reasoning behaviors.

This project fills the gap in reasoning model evaluation, emphasizes the importance of AI credibility and interpretability, provides researchers and developers with tools to address faithfulness issues, and promotes the industry's development toward more reliable and transparent AI systems.