Analysing Chain of Thought Dynamics: Active Guidance or Unfaithful Post-hoc Rationalisation?

TL;DR

This paper investigates the dynamics and faithfulness of Chain-of-Thought prompting in soft-reasoning tasks, revealing that its influence and faithfulness are often misaligned across different model types.

Contribution

It provides a comparative analysis of CoT's reliance and faithfulness in instruction-tuned, reasoning, and reasoning-distilled models, highlighting their differences.

Findings

CoT influence varies across model types

Faithfulness of CoT is not always aligned with its influence

Differences in reliance on CoT affect reasoning performance

Abstract

Recent work has demonstrated that Chain-of-Thought (CoT) often yields limited gains for soft-reasoning problems such as analytical and commonsense reasoning. CoT can also be unfaithful to a model's actual reasoning. We investigate the dynamics and faithfulness of CoT in soft-reasoning tasks across instruction-tuned, reasoning and reasoning-distilled models. Our findings reveal differences in how these models rely on CoT, and show that CoT influence and faithfulness are not always aligned.