Analysing Chain of Thought Dynamics: Active Guidance or Unfaithful Post-hoc Rationalisation?
This addresses the problem of unreliable reasoning in AI models for researchers and practitioners, but it is incremental as it builds on prior work on CoT limitations.
The paper investigated the dynamics and faithfulness of Chain-of-Thought (CoT) reasoning in soft-reasoning tasks, finding that CoT often provides limited gains and can be unfaithful to models' actual reasoning, with differences in reliance across model types.
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.