Reasoning Without Self-Doubt: More Efficient Chain-of-Thought Through Certainty Probing
Abstract
Recent large language models with chain-of-thought reasoning capabilities exhibit poor token efficiency due to Hesitation - spending excessive tokens verifying already-correct answers. Using our Probe-In-The-Middle technique to analyze model states during reasoning, we propose Dynasor-CoT, a certainty-based approach for dynamic reasoning termination. Our training-free method efficiently achieves up to 29% token reduction while maintaining accuracy across mathematical reasoning tasks like AMC23, AIME24, and MATH500.
Cite
Text
Fu et al. "Reasoning Without Self-Doubt: More Efficient Chain-of-Thought Through Certainty Probing." ICLR 2025 Workshops: FM-Wild, 2025.Markdown
[Fu et al. "Reasoning Without Self-Doubt: More Efficient Chain-of-Thought Through Certainty Probing." ICLR 2025 Workshops: FM-Wild, 2025.](https://mlanthology.org/iclrw/2025/fu2025iclrw-reasoning/)BibTeX
@inproceedings{fu2025iclrw-reasoning,
title = {{Reasoning Without Self-Doubt: More Efficient Chain-of-Thought Through Certainty Probing}},
author = {Fu, Yichao and Chen, Junda and Zhuang, Yonghao and Fu, Zheyu and Stoica, Ion and Zhang, Hao},
booktitle = {ICLR 2025 Workshops: FM-Wild},
year = {2025},
url = {https://mlanthology.org/iclrw/2025/fu2025iclrw-reasoning/}
}