Reproduce “Zero-Shot Chain-of-Thought Reasoning Guided by Evolutionary Algorithms in Large Language Models” 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
Feihu Jin, Yifan Liu, Ying Tan
Researchers from Peking University developed Zero-shot EoT (Evolutionary-guided Chain-of-Thought) prompting, a method that uses Large Language Models as evolutionary optimizers to dynamically generate and select instance-specific Chain-of-Thought prompts. The approach consistently outperforms static zero-shot CoT methods and competes with few-shot CoT methods on arithmetic and symbolic reasoning tasks without needing any examples.
Key insights
- A static, uniform Chain-of-Thought (CoT) prompt is suboptimal for all problem instances due to the diverse nature of reasoning problems and LLM pre-training dynamics.
- Large Language Models (LLMs) can function as effective evolutionary optimizers, performing operations like crossover and mutation to generate tailored CoT prompts.
- Rewriting the original problem statement guided by a selected CoT prompt significantly enhances the LLM's comprehension and subsequent reasoning performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.