A Gradient Perspective on RLVR Stability and Winner Advantage Policy Optimization
2026-06-15 • Machine Learning
Machine Learning
AI summaryⓘ
The authors studied how reinforcement learning methods that use verifiable rewards can sometimes become unstable when training language models. They analyzed how updating probabilities for the next word can either help or harm stability, depending on certain factors. Based on this, they created a new approach called Winner Advantage Policy Optimization (WAPO), which updates the model only when the feedback is clearly positive. Their method improved training stability and performed as well or better than existing methods on reasoning tasks.
reinforcement learninglanguage modelspolicy optimizationgradient dynamicsadvantage functionentropytoken probabilitymulti-hop question answeringtraining stabilitypolicy gradient
Authors
Prasanth YSS, Zhichen Ren, Rasa Hosseinzadeh, Ilan Gofman, Yuqi Chen, Zhaoyan Liu, Guangwei Yu, Jesse C. Cresswell, Satya Krishna Gorti
Abstract
Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but GRPO-style optimization remains prone to collapse. We analyse this instability through token-level gradient dynamics, deriving a taxonomy that predicts how updates affect next-token probabilities and entropy. The taxonomy shows that stability depends jointly on the advantage sign and token distribution under the current policy. Motivated by this finding, we propose Winner Advantage Policy Optimization (WAPO), a simple online clipped policy-gradient objective that updates only on positive-advantage completions. Across mathematical reasoning and multi-hop QA benchmarks, WAPO improves training stability and matches or outperforms baselines across multiple model families. Full code can be found at https://github.com/layer6ai-labs/wapo.