New ReAPO method boosts LLM reasoning via retry-aware credit assignment

Researchers have developed a new method called Retry-Aware Policy Optimization (ReAPO) to enhance the reasoning capabilities of large language models (LLMs). This approach addresses the issue of inefficient exploration in LLM reasoning by incorporating a novel credit assignment framework based on Stochastic Shortest Paths. ReAPO modifies existing GRPO algorithms to allow for more effective learning from failed attempts, reinforcing successful retries without inadvertently rewarding prior failures. Experiments on Qwen3 and OLMo-2 models across various benchmarks demonstrated that ReAPO improves first-attempt performance and achieves peak accuracy with significantly fewer generated tokens compared to other methods. AI

IMPACT Improves LLM reasoning efficiency and performance by optimizing exploration and credit assignment during training.

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.LG →

• GRPO
• LLM
• OLMo-2
• Qwen3
• RLVR
• Stochastic Shortest Paths

• paper
• model release

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