New LCMV method boosts LLM reasoning accuracy by conditioning on chain length

Researchers have introduced Length-Conditioned Majority Voting (LCMV), a novel approach to enhance the reasoning accuracy of large language models (LLMs). LCMV builds upon standard Majority Voting by incorporating an observation that longer chains of thought often correlate with lower correctness probabilities. This method models this bias using a logistic function and adapts it using unlabeled test data without requiring LLM fine-tuning. Experiments demonstrate that LCMV consistently outperforms existing aggregation strategies across various LLMs and reasoning benchmarks. AI

IMPACT This research offers a method to improve LLM reasoning accuracy without extensive retraining, potentially making LLMs more reliable for complex tasks.

RANK_REASON The item describes a new method presented in a research paper for improving LLM reasoning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Hugging Face Daily Papers →

• aimagelab/LCMV
• Chain-of-Thoughts
• GitHub
• Hugging Face
• large-language models
• Length-Conditioned Majority Voting
• logistic function
• Majority Voting

• paper
• model release

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