New ACPO framework enhances reinforcement learning for LLMs

Researchers have introduced Adaptive Credit Policy Optimization (ACPO), a new framework designed to improve credit assignment in reinforcement learning for large language models. ACPO addresses the challenge of sparse rewards by asymmetrically modulating policy updates, focusing on uncertain decisions in successful rollouts and overconfident tokens in failed ones. This method aims to maintain policy-gradient direction while improving performance on benchmarks like AIME 2025 and HumanEvalPro, outperforming existing methods such as DAPO, GTPO, and SAPO. AI

IMPACT This new ACPO framework could lead to more efficient training of LLMs, improving their performance on complex reasoning and coding tasks.

RANK_REASON The cluster contains a research paper detailing a new method for reinforcement learning in AI. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AIME 2025
• arXiv
• DAPO
• GTPO
• HumanEvalPro
• large language models
• reinforcement learning

• paper
• model release

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