TurnOPD: Making On-Policy Distillation Turn-Aware for Efficient Long-Horizon Agent Training

arXiv cs.AI Papers

Summary

TurnOPD introduces turn-level budgeting for on-policy distillation of long-horizon agents, addressing inefficiencies in vanilla OPD by adaptive rollout-depth and progressive turn-normalized loss budgeting, achieving better accuracy under equal training budgets.

arXiv:2607.05804v1 Announce Type: new Abstract: On-policy distillation (OPD) trains a student policy by matching a stronger teacher on the student's own trajectories, offering a promising framework for language agent training. However, its application to long-horizon agentic tasks remains insufficiently explored. We identify two key inefficiencies in vanilla agent OPD: (1) full-horizon rollouts often waste wall-clock resources on tail turns that provide weak and noisy KL supervision, and (2) trajectory-level KL objectives concentrate most of the loss on shallow tokens, leaving deeper decision turns under-trained once initial behaviors are aligned. To address these challenges, we propose TurnOPD, a turn-level budgeting strategy for efficient on-policy distillation of long-horizon agents. TurnOPD consists of two budget controllers: adaptive rollout-depth budgeting, which uses probe-based turn statistics to determine rollout length, and progressive turn-normalized loss budgeting, which gradually shifts KL weighting from token-level to turn-balanced supervision. Experiments on ALFWorld, WebShop, and Multi-Hop Search with task-specialized teacher models show that TurnOPD achieves superior validation accuracy under equal wall-clock training budgets and advances the accuracy--time frontier beyond vanilla OPD.
Original Article
View Cached Full Text

Cached at: 07/08/26, 04:38 AM

# TurnOPD: Making On-Policy Distillation Turn-Aware for Efficient Long-Horizon Agent Training
Source: [https://arxiv.org/abs/2607.05804](https://arxiv.org/abs/2607.05804)
[View PDF](https://arxiv.org/pdf/2607.05804)

> Abstract:On\-policy distillation \(OPD\) trains a student policy by matching a stronger teacher on the student's own trajectories, offering a promising framework for language agent training\. However, its application to long\-horizon agentic tasks remains insufficiently explored\. We identify two key inefficiencies in vanilla agent OPD: \(1\) full\-horizon rollouts often waste wall\-clock resources on tail turns that provide weak and noisy KL supervision, and \(2\) trajectory\-level KL objectives concentrate most of the loss on shallow tokens, leaving deeper decision turns under\-trained once initial behaviors are aligned\. To address these challenges, we propose TurnOPD, a turn\-level budgeting strategy for efficient on\-policy distillation of long\-horizon agents\. TurnOPD consists of two budget controllers: adaptive rollout\-depth budgeting, which uses probe\-based turn statistics to determine rollout length, and progressive turn\-normalized loss budgeting, which gradually shifts KL weighting from token\-level to turn\-balanced supervision\. Experiments on ALFWorld, WebShop, and Multi\-Hop Search with task\-specialized teacher models show that TurnOPD achieves superior validation accuracy under equal wall\-clock training budgets and advances the accuracy\-\-time frontier beyond vanilla OPD\.

## Submission history

From: Yuhang Zhou \[[view email](https://arxiv.org/show-email/dca82c47/2607.05804)\] **\[v1\]**Tue, 7 Jul 2026 03:56:35 UTC \(2,259 KB\)

Similar Articles

Learning to Foresee: Unveiling the Unlocking Efficiency of On-Policy Distillation

arXiv cs.CL

This paper investigates the parameter-level mechanisms behind the efficiency of On-Policy Distillation (OPD) for large language models, attributing it to early 'foresight' in module allocation and update direction. It proposes EffOPD, a plug-and-play method that accelerates OPD training by 3x without compromising final performance.

ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation

arXiv cs.LG

ShortOPD proposes a short-to-long on-policy distillation schedule that recovers pruned LLMs for free-form generation by focusing training on effective prefixes, achieving up to 9x improvement over unrecovered models and matching long-horizon distillation with a quarter of the training time.