What it is
Teacher-free on-policy distillation for flow matching that turns the model's own local SDE exploration into dense per-step reward supervision. Instead of distilling from a task-specific teacher, it distills from the student's own branched SDE rollouts: better-rewarded branches pull the velocity field, worse ones push it away, with reward-level fusion for multi-objective alignment.
Key results
- Beats prior RL methods on single-reward benchmarks
- Beats teacher-based OPD on mixed-reward benchmarks
- Needs no task-specific teacher while keeping dense per-step supervision
Dense on-policy supervision can be recovered from the student's own local exploration, avoiding teacher cost and teacher-student mismatch.