Reproduce “Rethinking Chain-of-Thought from the Perspective of Self-Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zongqian Wu, Baoduo Xu, Ruochen Cui, Mengmeng Zhan, Xiaofeng Zhu, Lei Feng
Researchers from UESTC and collaborating institutions developed a Chain-of-Thought (CoT) framework by reframing its iterative process as self-training, aiming to reduce prediction uncertainty. The approach achieves a +4.2% average accuracy improvement in zero-shot reasoning tasks and substantial computational efficiency gains by intelligently mitigating over-reasoning and promoting diverse thinking paths.
Key insights
- Chain-of-Thought (CoT) reasoning and self-training share the core objective of iteratively reducing prediction uncertainty.
- Semantic entropy can effectively quantify prediction uncertainty in CoT, correlating with accuracy and providing a metric for guiding iterative refinement.
- Addressing over-reasoning and promoting diversity in iterative CoT are crucial for sustained performance gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.