FutureBridge-OPD (FTB) validates teacher guidance through the future trajectories it induces in the student. At a state with high teacher-student disagreement, it executes a short teacher bridge and compares student continuations from the original and bridged states. The bridge is retained for distillation only when the resulting continuation is more strongly preferred by the teacher.

Publication

Preprint available as arXiv:2608.01953. The first version was submitted on August 3, 2026, and the latest listed revision (v2) was submitted on August 5, 2026.

Overview

In multi-turn agentic on-policy distillation, student errors can accumulate and move later states away from regions where teacher guidance is useful. High teacher-student disagreement identifies promising intervention points, but does not by itself show that an intervention will help. FutureBridge-OPD validates a candidate teacher bridge by examining its effect on subsequent student behavior before retaining it for distillation.

Evaluation

The paper evaluates FTB on ALFWorld, WebShop, and ScienceWorld. Under the main Qwen3-32B teacher to Qwen3-1.7B student setting, it reports average improvements of 16.6 and 7.6 points over vanilla OPD and TCOD, respectively, with additional experiments across student scales and teacher settings.

Resources