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

Name: TurnOPD: Making On-Policy Distillation Turn-Aware for Efficient Long-Horizon Agent Training
Uploaded: 2026-10-04T03:23:32.516215+00:00
Duration: 353 s
Description: TurnOPD improves long-horizon agent training by using a turn-level budgeting strategy for distillation. It addresses inefficiencies where vanilla methods waste compute on tail turns and concentrate loss on shallow tokens. By dynamically adjusting rollout depth and KL loss aggregation, TurnOPD achieves superior validation accuracy with less wall-clock time.

TurnOPD improves long-horizon agent training by using a turn-level budgeting strategy for distillation. It addresses inefficiencies where vanilla methods waste compute on tail turns and concentrate loss on shallow tokens. By dynamically adjusting rollout depth and KL loss aggregation, TurnOPD achieves superior validation accuracy with less wall-clock time.

Open the full episode