Reconciling Process Supervision with Outcome-Based Credit in Agentic Policy Optimization
TASPO uses privileged training signals only to redistribute verified outcome credit across agent actions.
The paper argues that process-style supervision can be too fine-grained, or misaligned with executable decisions, unless it is tied back to the task outcome. TASPO builds action-level weights from successful trajectories, keeping the verified outcome in charge of update direction and average scale. Across three agentic benchmarks, the authors report a 10.6% improvement over GRPO and better generalization to unseen tasks. ArXiv · AI/CL/LG's note
The paper argues that process-style supervision can be too fine-grained, or misaligned with executable decisions, unless it is tied back to the task outcome. TASPO builds action-level weights from successful trajectories, keeping the verified outcome in charge of update direction and average scale. Across three agentic benchmarks, the authors report a 10.6% improvement over GRPO and better generalization to unseen tasks. ArXiv · AI/CL/LG's note
score 5