Analysing Chain of Thought Dynamics: Active Guidance or Unfaithful Post-hoc Rationalisation?
XingweiT
Abstract
Investigation into Chain-of-Thought dynamics and faithfulness across various models reveals inconsistencies in their reliance on CoT and its alignment with actual reasoning.
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.
Community
Our paper investigates the faithfulness of CoT on the soft-reasoning tasks based on instruction-tuned, multi-step reasoning, and distilled reasoning models. We designed two experiments: 1) forcing an answer at intermediate reasoning steps to measure the gold answer confidence; 2) adding cues to mislead the models to measure the variance in the gold answer confidence. We found CoT often serves as a post-hoc justification for the instruction-tuned LLMs, but distilled reasoning LLMs rely heavily on CoT. Moreover, we found that unfaithful CoTs can still provide active guidance.
This work has been accepted by EMNLP 2025 (Main).
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Large Reasoning Models are not thinking straight: on the unreliability of thinking trajectories (2025)
- Can We Predict Alignment Before Models Finish Thinking? Towards Monitoring Misaligned Reasoning Models (2025)
- How Chain-of-Thought Works? Tracing Information Flow from Decoding, Projection, and Activation (2025)
- Thinking with Nothinking Calibration: A New In-Context Learning Paradigm in Reasoning Large Language Models (2025)
- Cognitive Chain-of-Thought: Structured Multimodal Reasoning about Social Situations (2025)
- Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens (2025)
- Logit Arithmetic Elicits Long Reasoning Capabilities Without Training (2025)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment:
@librarian-bot
recommend
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper