AI summaryⓘ
The authors address a problem in reinforcement learning where traditional methods assign equal credit to all actions based only on the final outcome, which can unfairly punish useful attempts or reward unhelpful actions. They introduce TRIAGE, a system that classifies actions into roles like progress or regression and assigns rewards accordingly, improving credit assignment. This approach reduces errors in learning signals and leads to better performance across several benchmark tasks. The main benefit comes from recognizing regression within successful attempts, while crediting exploration also helps. Overall, TRIAGE helps the learning agent make smarter decisions by understanding the role of each action better than before.
Reinforcement learningCredit assignmentGRPOTRIAGEPolicy gradientsSemantic rolesAdvantage estimationExplorationRegressionOutcome rewards
Authors
Yuanda Xu, Zhengze Zhou, Hejian Sang, Xiaomin Li, Jiaxin Zhang, Xinchen Du, Zhipeng Wang, Alborz Geramifard
Abstract
Agentic reinforcement learning requires assigning credit to environment-facing actions such as searches, clicks, edits, navigation commands, and object interactions. Standard GRPO uses the final verifier outcome as a uniform advantage over all action tokens. This outcome signal is useful but structurally incomplete: it punishes useful exploration in failed rollouts and reinforces redundant or regressive actions in successful rollouts. We propose TRIAGE, a role-typed credit assignment framework that adds a semantic role axis to outcome credit. A structured judge classifies each segment as decisive progress, useful exploration, no-progress infrastructure, or regression, and a fixed role-conditioned rule maps these labels to bounded segment-level process rewards. This keeps verifier outcomes as the source of optimization direction while correcting the two main blind spots of outcome-only credit. We further show that role-conditioned credit is the optimal segment-level correction expressible from role labels alone -- a projection of the per-segment advantage residual onto the role variable -- so that the fixed role constants reduce advantage estimation error whenever the judge is reliable, and we connect this to lower-variance policy gradients. Across ALFWorld, Search-QA, and WebShop, TRIAGE improves success rates over GRPO for two policy models and outperforms both a scalar judge-derived process reward and an outcome-supervised shared-backbone value baseline. Ablations show that the gain comes from role typing rather than merely adding dense rewards: reliable detection of regression inside successful trajectories is the dominant contributor, while exploration credit provides a consistent secondary gain; on completed ALFWorld and WebShop rollouts, TRIAGE also reduces environment-facing turns by an additional $10.4\%$ and $14.8\%$ relative to GRPO.