[2609.38972] Making LLMs Say What They Think: Measuring and Improving CoT-Interpretability Alignment

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 this https URL.

View PDF HTML (experimental) 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…

saved by

• Ishaan Panigrahi

related reading

• [2510.24941] Can Aha Moments Be Fake? Towards Quantifying Decorative and True Thinking in Chain-of-Thoughtarxiv.org
• Transformers Provably Learn to Internalize Chain-of-Thoughtarxiv.org
• Quantifying the Necessity of Chain of Thought through Opaque Serial Deptharxiv.org
• [2412.06769] Training Large Language Models to Reason in a Continuous Latent Spacearxiv.org
• Measuring Faithfulness in Chain-of-Thought Reasoning \ Anthropicanthropic.com
• [2201.11903] Chain of Thought Prompting Elicits Reasoning in Large Language Modelsarxiv.org
• [2405.18915] Towards Faithful Chain-of-Thought: Large Language Models are Bridging Reasonersar5iv.labs.arxiv.org
• Towards Faithful Chain-of-Thought: Large Language Models are Bridging Reasonersarxiv.org
• Training Large Language Models to Reason in a Continuous Latent Spacearxiv.org
• [2507.11473] Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safetyarxiv.org
• the case for CoT unfaithfulness is overstated — LessWronglesswrong.com
• [2510.09714] All Code, No Thought: Current Language Models Struggle to Reason in Ciphered Languagearxiv.org