STEP: Success-Rate-Aware Trajectory-Efficient Policy Optimization
This addresses the problem of inefficient and unstable training in multi-turn online reinforcement learning for researchers and practitioners, but it is incremental as it builds on trajectory-level optimization methods.
The paper tackled the inefficiency and misleading learning signals in multi-turn online reinforcement learning by proposing STEP, a framework that dynamically allocates sampling based on per-task success rates and performs step-level optimization, resulting in improved sample efficiency and training stability, with experiments showing faster convergence and better generalization under the same sampling budget.
Multi-turn interaction remains challenging for online reinforcement learning. A common solution is trajectory-level optimization, which treats each trajectory as a single training sample. However, this approach can be inefficient and yield misleading learning signals: it applies uniform sampling across tasks regardless of difficulty, penalizes correct intermediate actions in failed trajectories, and incurs high sample-collection costs. To address these issues, we propose STEP (Success-rate-aware Trajectory-Efficient Policy optimization), a framework that dynamically allocates sampling based on per-task success rates and performs step-level optimization. STEP maintains a smoothed success-rate record to guide adaptive trajectory resampling, allocating more effort to harder tasks. It then computes success-rate-weighted advantages and decomposes trajectories into step-level samples. Finally, it applies a step-level GRPO augmentation to refine updates for low-success tasks. Experiments on OSWorld and AndroidWorld show that STEP substantially improves sample efficiency and training stability over trajectory-level GRPO, converging faster and generalizing better under the same sampling budget.