Self-OPD is a teacher-free on-policy distillation framework for flow matching models that replaces a separate, task-specific teacher with the student model’s own self-exploration for step-wise supervision. At each timestep it branches the next-state prediction into several stochastic candidates, rolls them out, and compares their rewards against a deterministic self-reference baseline to compute training signal, avoiding both the cost of training per-task teachers and the compounding errors that arise from teacher-student distribution mismatch. On single and mixed-reward benchmarks, Self-OPD outperforms prior reinforcement learning and on-policy distillation methods that rely on dedicated teacher models.
