AGPO: Asymmetric Group Policy Optimization for Verifiable Reasoning and Search Ads Relevance at JD

A new reinforcement learning method, Asymmetric Group Policy Optimization (AGPO), is proposed to improve reasoning in large language models (LLMs) while preserving exploration capacity. Current RLVR methods enhance sampling efficiency but narrow reasoning boundaries compared to base models. AGPO uses a negative-dominant strategy to suppress incorrect paths and a group advantage mechanism for positive updates, focusing on rare correct paths. The paper is published on arXiv under ID 2605.05826, with applications at JD.com for search ads relevance.