Get Started
Home
Topics
Search
Library
Research questionHow reliably can chain-of-thought text reveal which reasoning steps causally drive correct answers?A reasoning step can look coherent without being functionally necessary for the final answer. The difficulty is determining whether the text of a step reveals its causal contribution to success.
AI
Alignment & Safety
Evaluation & Benchmarks
Mechanistic Interpretability
Natural Language Processing
Reasoning
Latest papersRecent research connected to this question, newest first.Legibility is Not Interpretability: Comparing Judged and Actual Importance in Chain-Of-Thought ReasoningThe evidence concerns language-model chain-of-thought traces and defines step importance by the change in expected reward, such as final-answer correctness, when the step is included, estimated through Monte Carlo rollouts. It evaluates LLM judges and a fine-tuned step-level critic, finding that importance is only partially recoverable from trace text, with different performance for incorrect and correct responses.research paper · Sep 3, 2026
Related questions
Can chain-of-thought monitoring detect consequential computation hidden in semantically irrelevant filler tokens?Can chain-of-thought monitoring detect preferences received through tools or inferred from raw artifacts?How can multimodal chain-of-thought be compressed without losing image-dependent reasoning or inducing hallucinations?How can industrial root-cause analysis reconcile conflicting free-text hypotheses without slow, deeply agentic LLM workflows?