ATOD: Annealed Turn-aware On-policy Distillation for Multi-turn Autonomous Agents

Researchers propose ATOD, a hybrid online distillation algorithm that combines on-policy distillation and reinforcement learning to train small language-model agents for multi-turn tasks. ATOD uses an annealed schedule to leverage OPD's early efficiency and RL's exploratory improvement, achieving a 3.03-point average success rate improvement over OPD and surpassing teacher models by 2.16 points on ALFWorld, WebShop, and Search-QA.

arXiv:2606.27814v1 Announce Type: new Abstract: Training small language-model agents for long-horizon interactive tasks requires both fast imitation and reward-driven improvement. On-policy distillation OPD provides dense teacher guidance and typically improves rapidly in the early stage, but its gains saturate once the student approaches the teacher, limiting the final performance ceiling. Reinforcement learning RL directly optimizes environment rewards and encourages exploratory improvement toward a higher reward-defined ceiling, but sparse and delayed feedback makes early-stage learning much less efficient than OPD. In this paper, we propose ATOD Annealed Turn-aware On-policy Distillation , a hybrid online distillation algorithm that explicitly exploits this complementarity. 1 ATOD uses an annealed OPD-RL schedule: OPD dominates early training to approach teacher-level behavior, while RL is gradually strengthened to drive reward-based exploration. 2 ATOD introduces Turn-level Disagreement-Uncertainty Reweighting T-DUR , which softly amplifies high-utility turns and improves dense supervision in long trajectories. Experiments on ALFWorld, WebShop, and Search-QA show that ATOD consistently outperforms competing post-training baselines: across the three student sizes, ATOD improves average success rate by 3.03 points over OPD and 23.62 points over GRPO, while surpassing the corresponding teacher models by 2.16 points.