Making LLMs Say What They Think: Measuring and Improving CoT-Interpretability Alignment

2026-09-30

AI Summary

This study addresses the inconsistency between Chain-of-Thought (CoT) reasoning in large language models and their internal computational processes, which undermines the credibility of explicit reasoning. To tackle this issue, we propose the CIA metric, introducing the first mechanistic interpretability-based framework for quantifying CoT alignment. Building upon this metric, we further develop a faithfulness-grounded reward mechanism to optimize model behavior through reinforcement learning during post-training. This work establishes the first closed-loop pipeline for systematically measuring and optimizing CoT-interpretability alignment. Our approach significantly enhances the parametric faithfulness of generated reasoning while preserving or even improving accuracy on downstream tasks.

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.

Problem

• Chain-of-Thought
• Interpretability
• Faithfulness
• Large Language Models
• Alignment

Innovation

• Chain-of-Thought
• Interpretability
• Parametric Faithfulness
• Post-training
• Large Language Models