LLMs often fail to reveal true reasoning in chain of thought
Making LLMs Say What They Think: Measuring and Improving CoT-Interpretability Alignment
Computation and LanguageArtificial Intelligence
Summary
Large Language Models (LLMs) often write out their reasoning steps, called chain-of-thought (CoT), to explain how they get answers. However, these explanations don’t always match how the model actually thinks internally. The authors introduce a new way to measure how well the CoT matches the model’s true reasoning and show that current models have limited alignment. They also try training models to make their written reasoning closer to their real computations without hurting accuracy. This work helps make LLM reasoning more trustworthy by auditing and improving their explanations.
What this means in practice
- •For machine learning engineers: Improve model training to generate reasoning steps that better reflect internal computations while maintaining accuracy.
- •For ai system auditors: Use the CoT-Interpretability Alignment metric to evaluate if LLM explanations faithfully represent underlying model reasoning.
Authors
Yihuai Hong, Shauli Ravfogel, Chen Zhao, Eunsol Choi
Abstract
Chain-of-thought (CoT) traces often serve as a proxy for how Large Language Models (LLMs) arrive at their answers. However, growing evidence shows that models' CoT often fails to reflect their internal computations and can be changed without affecting their final answers. In this work, we measure and improve the alignment between the reasoning described in an LLM's CoT and what it computes internally. We propose CoT-Interpretability Alignment (CIA), a metric that measures the agreement between a model's CoT traces and its internal reasoning strategies as detected by interpretability tools. We evaluate CIA on three tasks (two-hop question answering, hint intervention, and integer multiplication) across three LLMs, finding that LLMs exhibit limited alignment across all tasks (44.8-75.9%). We then experiment with improving CIA via post-training, setting both the task accuracy and parametric faithfulness signals as a reward. Experiments show that we can substantially improve CoT parametric faithfulness while maintaining or improving the task accuracy. We provide rich analysis, such as their generalization patterns. Our work provides both a framework for auditing CoT parametric faithfulness and a pathway toward making models' explicit reasoning more trustworthy. Code and data are available at https://github.com/yihuaihong/CIA-minimal-repro.