Agentic-OPD: Recovering Forgotten Agent Capabilities via On-Policy Distillation for Agentic Tasks
Saved training checkpoints recovered lost agent capabilities, improving results by up to 7.4 points in the authors’ tests.
The researchers found that the best-validation checkpoint failed on problems earlier checkpoints from the same training run had solved. Agentic-OPD selects those saved checkpoints as teachers according to the student’s remaining gaps. It limits distillation to failed student attempts on problems where the teacher performs better, aiming to preserve already-mastered behavior. Across five task domains and 24 benchmarks and task subsets, the authors report gains including a 4B student outperforming GPT-5.6 and DeepSeek-V4.1 on deep search.
ArXiv · AI/CL/LG's note
The researchers found that the best-validation checkpoint failed on problems earlier checkpoints from the same training run had solved. Agentic-OPD selects those saved checkpoints as teachers according to the student’s remaining gaps. It limits distillation to failed student attempts on problems where the teacher performs better, aiming to preserve already-mastered behavior. Across five task domains and 24 benchmarks and task subsets, the authors report gains including a 4B student outperforming GPT-5.6 and DeepSeek-V4.1 on deep search.
ArXiv · AI/CL/LG's note
score 5