New MPPA method boosts LLM reasoning accuracy with diverse plans

Researchers have developed a new method called Multi-Path Plan Aggregation (MPPA) to enhance the reasoning capabilities of large language models during inference. MPPA utilizes separate LoRA adapters on a shared backbone to generate multiple diverse plans and their corresponding solutions. An aggregator then scores these solutions, leveraging reasoning evidence to select the most accurate final answer. This approach improves the accuracy-token trade-off compared to traditional methods, achieving higher accuracy with fewer tokens. AI

IMPACT This method could lead to more efficient and accurate LLM reasoning, particularly in complex tasks like mathematics and science.

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

Read on arXiv cs.CL →

• arXiv
• Hugging Face
• LoRA
• Multi-Path Plan Aggregation
• Siheng Xiong

• paper
• model release

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