The paper introduces Task-Progress Distillation (TPD), an offline method that pairs each action in teacher demonstrations with a short label indicating the current task stage. A small 1.7B student model trained with 404 demonstrations using TPD achieves 72.4% mean unseen task success on ALFWorld, matching action-only supervision and outperforming a reasoning-trained student. At lower demonstration budgets, explicit stage labels significantly boost performance, raising success from 48.0% to 67.7% with 200 demos, and both approaches converge to 76.9% with 808 demos.
arXiv:2608.22364v1 Announce Type: new Abstract: World action models (WAMs) couple visual future prediction with robot action generation, but accelerated students can lose task capabilities during dis...
State-OPD introduces a state‑grounded on‑policy distillation framework that separates the objective task and action‑observation record from the student’s internal reasoning trajectory. By removing prior student reasoning and resampling complete reasoning‑action responses at high‑entropy states, the method improves teacher supervision and action diversity. Evaluations on ALFWorld, ScienceWorld, and WebShop show State‑OPD achieves the highest macro‑average success rates for Qwen3‑1.7B and Qwen3‑4B students, even surpassing a larger teacher model on certain tasks.
arXiv:2609.36608v1 Announce Type: new Abstract: On-policy distillation (OPD) trains multi-turn language agents with dense teacher supervision on student-generated responses. However, standard think-t...
arXiv:2606. 27814v4 Announce Type: replace Abstract: Training small language-model agents for long-horizon interactive tasks requires both fast imitation and reward-driven improvement.
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.
arXiv:2606. 10385v1 Announce Type: cross Abstract: On-policy distillation (OPD) has demonstrated strong empirical gains in enhancing complex reasoning in LLMs by aligning a student model with a teacher's predictive distribution over the student's own trajectories.
Agentic Critical Training (ACT) is a reinforcement learning approach that trains language‑model agents to judge actions directly by pairing expert actions with alternatives and rewarding correct selection. Unlike imitation learning or self‑reflection, ACT does not require reference rationales and can be reused across model sizes. Experiments on Qwen3‑8B and Olmo‑3‑7B‑Instruct show that ACT improves performance on ALFWorld, WebShop, ScienceWorld, and outperforms supervised reflection baselines, with significant gains over IL and IL→RL baselines.
arXiv:2606. 03461v1 Announce Type: new Abstract: Stronger code agents are commonly assumed to be superior teachers for post-training, yet this assumption remains poorly disentangled from task difficulty, harness design, and student capacity.
arXiv:2509. 14257v3 Announce Type: replace-cross Abstract: Large Language Model agents achieve strong performance on multi-step reasoning and tool-use tasks, but their impressive capabilities typically rely on extremely large backbones.
RoboCoach introduces a world-model-guided coaching framework that uses imagined failures to direct demonstration requests and expert updates for robot manipulation tasks. The system, called RIDI, runs reusable skill experts within a shared action-conditioned world model and records the first failing subtask to decide which demonstrations to acquire and which adapters to refine. Experiments on simulation suites and real robots show that with only 150 additional subtask demonstrations, success rates increase dramatically, and the coached experts transfer effectively to unseen task compositions.
TAP (Trajectory‑Anchored Pruning) is a structural pruning framework designed for reinforcement‑learning‑trained agents that addresses the shortcomings of existing pruning methods on long‑horizon tasks. It couples iterative channel pruning with on‑policy recovery, anchoring student interactions to teacher trajectories and re‑scoring channels using gradients from the recovery objective. Experiments show that removing 60% of FFN channels preserves 99.2% and 88.0% of task success rates on ALFWorld and WebShop while cutting GPU time per successful task by about 22% and 17%.
WAM-OPD introduces a method for post‑training improvement of World Action Models (WAMs) by collecting student rollout histories and querying a stronger teacher for paired video and action targets. The student learns from both modalities while preserving its one‑step generation capability at deployment. Experiments on 12 RoboTwin 2.0 tasks and four real‑robot tasks show success rates rising from 33.8% to 65.7% and from 51.4% to 64.6%, respectively, with joint video‑action supervision outperforming single‑modality approaches.