AIAny
Icon for item

AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning

Converts sparse trajectory-level rewards into turn-level credit by aggregating token-level teacher–student log-probability gaps and recursively updating a Bayesian belief in log-odds; produces turn-wise reweighting for policy optimization without an extra critic or rollouts.

Introduction

Long-horizon, multi-turn agentic tasks hide which few decisions actually determine outcomes because verifiable rewards are sparse and delayed. AgentOPSD turns sparse outcome supervision into interpretable turn-level credit by aggregating token-level teacher–student log-probability gaps into turn evidence and recursively updating a Bayesian belief state in log-odds, allowing standard policy updates to spotlight pivotal turns without adding a critic or extra rollouts.

Key Findings
  • Recursive Bayesian aggregation: token-level teacher–student log-prob gaps are summed into turn evidence and used to update a belief state in log-odds; the marginal belief revision between consecutive states identifies pivotal turns.
  • Critic-free integration: the method yields turn-level advantage reweighting that plugs into standard policy optimization (e.g., GRPO) without requiring an auxiliary critic or additional environment interactions.
  • Empirical improvements: outperforms GRPO and strong self-distillation baselines on ALFWorld, WebShop, and Search-QA; reported 89.1% success on ALFWorld with Qwen2.5-7B. Ablations attribute gains to turn-level aggregation and history-dependent recursion.
  • Practical trade-offs: uses teacher–student log-probabilities as supervision, so it leverages privileged teacher signals rather than changing rollout dynamics.
Who it's for & trade-offs

Great fit if you train LLM-based agents on long-horizon, multi-turn tasks where final outcomes are verifiable but intermediate credit is unclear, and you can evaluate a teacher policy to obtain token-level log-probabilities. It is especially relevant when you want to keep rollout behavior unchanged and avoid learning a separate critic.

Look elsewhere if you cannot run a teacher model to supply log-prob gaps, if environment rewards are already dense and informative, or if you need methods that explicitly learn value functions for off-policy correction. Computational overhead is mainly teacher evaluations and bookkeeping of recursive belief updates; no extra environment samples are required.

Where it fits

Compared with contemporaneous self-distillation approaches, AgentOPSD emphasizes a recursive, history-dependent Bayesian belief in log-odds to produce turn-level marginal revisions. That contrasts with methods that localize supervision to action spans or apply alternate Bayes-calibrated scores: AgentOPSD’s recursion explicitly models how each turn changes posterior belief about the final outcome, yielding principled reweighting for policy updates.

Information

  • Websitearxiv.org
  • AuthorsZi-Han Wang, Zhengxi Lu, Zhiyuan Yao, Jinyang Wu, Jie Wu, Zhengzhou Cai, Yueqing Sun, Ziang Ye, Linji Hao, Qi Gu
  • Published date2026/08/06

More Items

Replaces external environment interaction in agentic RL training with 'world rehearsal': the policy alternates between making tool calls and simulating their environment responses, jointly optimizing both roles so the agent internalizes environment dynamics and improves long-horizon tool use and transfer.

Evaluates vision-language model judges on computer-using agent (CUA) trajectories to measure verifier reliability. Provides OSReward-Hard and OSReward-Multi challenge sets, the OS-Shepherd-100K reasoning-annotated corpus, and trained OS-Shepherd reward models that match commercial judges at ~30–60× lower cost.

Generates large-scale, explorable 3D open-world scenes from open-ended text prompts, producing editable instance-level assets and a consistent global terrain. Uses agentic planning to convert text into region/terrain/asset specifications and a coarse-to-fine pipeline for terrain construction, mesh reconstruction, and render-based refinement.