Stabilizing Rubric Integration Training via Decoupled Advantage Normalization
This addresses reward design problems in AI reasoning systems, offering a method to stabilize training and enhance reasoning quality, though it is incremental as it builds on existing policy optimization frameworks.
The paper tackled the limitations of outcome and process reward models in reinforcement learning for reasoning tasks by proposing Process-Aware Policy Optimization (PAPO), which integrates process-level evaluation via decoupled advantage normalization, resulting in improved performance such as 51.3% vs. 46.3% on OlympiadBench.
We propose Process-Aware Policy Optimization (PAPO), a method that integrates process-level evaluation into Group Relative Policy Optimization (GRPO) through decoupled advantage normalization, to address two limitations of existing reward designs. Outcome reward models (ORM) evaluate only final-answer correctness, treating all correct responses identically regardless of reasoning quality, and gradually lose the advantage signal as groups become uniformly correct. Process reward models (PRM) offer richer supervision, but directly using PRM scores causes reward hacking, where models exploit verbosity to inflate scores while accuracy collapses. PAPO resolves both by composing the advantage from an outcome component Aout, derived from ORM and normalized over all responses, and a process component Aproc, derived from a rubric-based PRM and normalized exclusively among correct responses. This decoupled design ensures that Aout anchors training on correctness while Aproc differentiates reasoning quality without distorting the outcome signal. Experiments across multiple model scales and six benchmarks demonstrate that PAPO consistently outperforms ORM, reaching 51.3% vs.\ 46.3% on OlympiadBench while continuing to improve as ORM plateaus and declines.