Self-OPD removes the teacher from on-policy distillation by branching each step into K stochastic candidates

arXiv 2608.26872, via Hugging Face Daily Papers·low signalXBlueskyLinkedInCopy link

Submitted 2026-08-27 with 66 upvotes, Self-OPD attacks two costs in on-policy distillation for flow matching models: training a task-specific teacher for every new objective, and compounding trajectory errors from teacher-student distribution mismatch. At each timestep it branches the deterministic next-state prediction into K stochastic SDE candidates, rolls them out with the ODE sampler, and compares their rewards to derive step-wise supervision from the student's own exploration. The teacher-free framing is the interesting economic claim: it turns a per-objective training cost into an inference-time search cost.