New method improves LLM reasoning by conditioning votes on chain length

Researchers have introduced Length-Conditioned Majority Voting (LCMV), a new method to enhance the reasoning accuracy of large language models (LLMs). LCMV leverages the observation that longer Chain-of-Thoughts (CoTs) are often less accurate. The approach models this as a length-conditioned prior probability, which is adapted at test time using unlabeled data without requiring LLM fine-tuning. This adapted prior then weights the votes of sampled answers, consistently outperforming standard majority voting and other aggregation strategies across various LLMs and reasoning benchmarks. AI

IMPACT Enhances LLM reasoning capabilities by optimizing vote aggregation based on chain length, potentially improving performance on complex tasks.

RANK_REASON The cluster contains an academic paper detailing a new method for improving LLM reasoning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Chain-of-Thoughts
• Fiorenzo Parascandolo
• Hugging Face
• large-language models
• Length-Conditioned Majority Voting
• Majority Voting

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →