RODS: Reward-Driven Online Data Synthesis for Multi-Turn Tool-Use Agents

2026-06-17

AI Summary

This work addresses the inefficiency in multi-turn tool-use reinforcement learning caused by the rapid exhaustion of high-information samples in static datasets. To overcome this limitation, the authors propose a closed-loop training framework that dynamically probes the policy’s capability boundary via reward variance, synthesizes new trajectories with matched structural complexity, and maintains a dynamic replay buffer that co-evolves with the policy. The approach incorporates a zero-overhead capability-boundary detection mechanism and a skill-aligned resampling strategy to enable efficient online data generation. Remarkably, it achieves performance comparable to offline methods using 17K samples with only approximately 800 active samples and 400 human seed demonstrations—reducing trajectory usage by about 20× and significantly outperforming fixed-data RL and environment-augmentation baselines.

Abstract

Multi-turn tool-use RL is bottlenecked by the rapid depletion of informative samples in static datasets. We observe that the gradient signal in GRPO concentrates on tasks with the highest rollout reward variance, a consequence of the Popoviciu upper bound. Consequently, samples near the agent's capability boundary -- where successes and failures are roughly balanced -- contribute disproportionately large policy gradients. As training progresses, this boundary continuously shifts, which gradually depletes the pool of informative samples in a static dataset. We propose RODS (Reward-driven Online Data Synthesis) to resolve this depletion. RODS closes the loop between RL training and data generation by repurposing the progress reward variance as a practical, zero-cost boundary detector that requires no extra inference beyond the rollouts already computed for training. It continuously identifies such boundary samples, synthesizes new multi-turn variants matching their structural complexity (e.g., API topology and dependency depth) via a skill-aligned resampling pipeline, and manages a dynamic replay buffer that co-evolves with the policy. Starting from 400 human seeds and maintaining an active training pool of ~800 samples, RODS achieves comparable performance to a 17K-sample offline pipeline while requiring roughly 20x fewer trajectories, and improves over fixed-data RL and environment augmentation in our controlled setting.

Problem

• multi-turn tool-use
• informative sample depletion
• static dataset
• reward variance
• capability boundary

Innovation

• Reward-Driven Synthesis
• Online Data Generation
• Multi-Turn Tool-Use
• Policy Gradient Variance
• Dynamic Replay Buffer