Selective Rollout: Mid-Trajectory Termination for Multi-Sample Agent RL

2026-05-07

AI Summary

This work addresses the inefficiency in multi-sample reinforcement learning, where substantial computational resources are wasted on redundant trajectories yielding identical final rewards. To mitigate this, the authors propose a dynamic termination mechanism that identifies zero-variance trajectory groups by measuring the edit distance between action sequence prefixes. By integrating Group Relative Policy Optimization (GRPO) with a single-parameter gating mechanism, the method terminates uninformative trajectories early during execution, thereby avoiding unnecessary computation. Evaluated on the ALFWorld benchmark using the Qwen2.5-7B model, the approach achieves a 10.7% acceleration in training and improves task success rates by 2.5 percentage points across 50 previously unseen tasks, demonstrating significant gains in both efficiency and generalization.

Abstract

Group-relative RL training (GRPO) samples a small group of parallel rollouts for every training prompt and uses their within-group reward spread to compute per-trajectory advantages. In agentic environments each rollout is a long multi-turn dialogue with one LLM call per step, so this multi-sample multiplier dominates the total training cost. When every rollout of a prompt ends with the same reward, the group has zero reward variance and contributes no gradient, so the extra rollouts add no information; such groups are common in practice (typically around 40% of all groups), so the wasted-compute fraction is substantial rather than marginal. Existing methods filter such groups at the prompt level, either after their rollouts are paid for or before any rollout begins, but both decide without using information that becomes available during the rollout itself. We instead ask whether the in-group divergence between the partial trajectories at an intermediate step can already predict that the group will be zero-variance: when the parallel rollouts have already converged on the same action prefix, the group is on track to produce a single reward, and we can stop early. We propose a one-parameter gate that stops a group when the mean pairwise prefix edit distance between its partial action sequences falls below a threshold. On a 60-iteration on-policy GRPO run on ALFWorld with Qwen2.5-7B, averaged over four random seeds, the gated arm finishes 10.7% faster in wall-clock (bootstrap 95% CI excludes 0) and shifts held-out success rate on 50 unseen tasks by +2.5 pp, with the held-out gain tracing to a measurable reduction in zero-advantage gradient-batch dilution. Code is available at https://github.com/zhiyuanZhai20/selective-rollout.

Problem

• multi-sample RL
• trajectory termination
• zero-reward variance
• compute efficiency
• agent training

Innovation

• Selective Rollout
• Mid-Trajectory Termination
• Group-relative Policy Optimization
• Prefix Edit Distance
• Zero-Variance Filtering