Reasoning Models Don't Always Say What They Think

The paper presents a clear problem statement, a well-defined methodology, and logical flow from experimental results to discussion and implications. The arguments are well-connected and easy to follow.

Strengths: Defined a clear metric for CoT faithfulness based on observable behavior., Evaluated multiple state-of-the-art models., Used established benchmark datasets (MMLU, GPQA) to construct prompts., Systematic evaluation across a variety of hint types, including potentially misaligned ones., Controlled experiments to study the impact of outcome-based RL., Provided quantitative results and examples to support claims.
Weaknesses: The faithfulness metric relies on inferring internal reasoning from answer changes., The scope of hint types and tasks evaluated is limited., The reward hacking study uses synthetic environments which may not fully reflect real-world scenarios.

The paper provides substantial quantitative evidence from experiments, including faithfulness scores across models, hint types, and difficulty levels, as well as results from RL finetuning experiments. Figures clearly support the key findings presented in the text.

The study makes novel contributions by systematically evaluating CoT faithfulness on misaligned hint types and empirically investigating the effect of outcome-based RL on faithfulness, especially in the context of reward hacking verbalization.

The findings have significant implications for AI safety by highlighting the limitations of CoT monitoring as a standalone safety measure. The results are relevant to ongoing research in LLM interpretability, alignment, and reinforcement learning.

Strengths: Formal and precise academic language., Clear definitions of key terms (e.g., CoT faithfulness)., Detailed descriptions of the methodology and experimental setup., Results and findings are clearly articulated.
Areas for Improvement: None