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

2025-08-27

AI Summary

This study challenges the dynamic guidance fidelity of chain-of-thought (CoT) reasoning in soft reasoning tasks—such as analytical and commonsense reasoning—questioning whether CoT reflects genuine internal reasoning or merely post-hoc, unfaithful rationalization. Method: Through dynamic behavioral analysis and attribution-based evaluation, we systematically compare CoT usage across instruction-tuned, dedicated reasoning, and distillation-based large language models. Contribution/Results: We find that CoT yields limited practical improvement for soft reasoning and frequently diverges from the model’s actual inference process. Moreover, the extent of CoT influence and user trust exhibit significant misalignment across model families. To our knowledge, this is the first work to empirically identify a pervasive “guidance failure” phenomenon of CoT in soft reasoning, directly challenging its default assumption as a reliable reasoning mechanism. Our findings provide novel empirical evidence and methodological insights for interpretable AI and trustworthy reasoning evaluation.

Technology Category

• Reasoning under Uncertainty: Causality
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Knowledge Representation and Reasoning: Reasoning with Beliefs

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

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.

Problem

• Investigating Chain-of-Thought dynamics in soft-reasoning tasks
• Assessing faithfulness of CoT to model's actual reasoning
• Comparing CoT reliance across different instruction-tuned models

Innovation

• Investigating Chain-of-Thought dynamics in reasoning models
• Analyzing faithfulness of CoT in soft-reasoning tasks
• Comparing instruction-tuned and reasoning-distilled model performance