Reproduce “C3oT: Generating Shorter Chain-of-Thought without Compromising Effectiveness” 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
Yu Kang, Xianghui Sun, Liangyu Chen, Wei Zou
This paper introduces a method to generate shorter Chain-of-Thought (CoT) reasoning without losing effectiveness
Key insights
- Using distinct initial prompt tokens to condition longer vs shorter CoT helps LLMs learn their relationships
- Mixed training with CoT at various compression levels leads to mutually beneficial effects
- Models can automatically select appropriate CoT length based on question difficulty
- Compression effectiveness depends on having sufficient key information preserved
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.