Understanding Chain-of-Thought in LLMs through Information Theory

TL;DR

This paper introduces an information-theoretic framework to evaluate Chain-of-Thought reasoning in LLMs, accurately identifying failure modes without relying on annotated data, and demonstrating superior performance on multiple datasets.

Contribution

It formalizes CoT reasoning using information theory, enabling precise failure mode detection without annotated datasets, and improves evaluation accuracy over existing methods.

Findings

Outperforms outcome-based methods in accuracy

Effectively identifies failure modes in reasoning

Validated on arithmetic and reasoning datasets

Abstract

Large Language Models (LLMs) have shown impressive performance in complex reasoning tasks through the use of Chain-of-Thought (CoT) reasoning, allowing models to break down problems into manageable sub-tasks. However, existing CoT evaluation techniques either require annotated CoT data or fall short in accurately assessing intermediate reasoning steps, leading to high rates of false positives. In this paper, we formalize CoT reasoning in LLMs through an information-theoretic lens. Specifically, our framework quantifies the `information-gain' at each reasoning step, enabling the identification of failure modes in LLMs without the need for expensive annotated datasets. We demonstrate the efficacy of our approach through extensive experiments on toy arithmetic, GSM8K and PRM800k datasets, where it significantly outperforms existing outcome-based methods by providing more accurate insights into model performance on individual subtasks.