Peking and Kling Team's Flux-OPD Tops OPD Baselines on Open Tasks

Most on-policy distillation work assumes a fixed teacher signal, which is fine when a task has verifiable rewards but wobbly when the target is something like rewriting a prompt for a video generator or answering an open clinical question. A new paper from Peking University and Kuaishou's Kling team, posted to Hugging Face with collaborators at Tsinghua, Shanghai Jiao Tong and Zhongguancun Academy, argues that the fix is to let the context the teacher sees evolve as the student improves, and then to be very careful about how that evolving signal is combined.

The method, Flux-OPD, decomposes the reverse-KL distillation objective into two pieces: a distillation term that pulls the student toward the normalized geometric mean of context-conditioned teacher distributions, and a conflict term that measures how much those context-conditioned teachers disagree. The practical move is to anchor training to the stable context-free teacher and inject the difference between the context-conditioned and context-free teachers as a correction, weighted down when the conflict term says the contexts are pulling in opposite directions. Contexts are re-extracted from student rollouts every 300 steps, which is what makes the supervision "evolving" rather than fixed.

The results the authors report are modest in absolute terms but consistent. On HealthBench with a Qwen3 8B teacher and 1.7B student, Flux-OPD scores 20.61 total, versus 19.63 for standard OPD, 19.66 for OEL, and 19.06 for the SFT-initialized student. On prompt optimization for video generation, evaluated on VBench and Video-Bench with Wan2.1-VACE-1.3B and CogVideoX-2B as downstream generators, plain OPD actually degrades the student on some settings, while Flux-OPD comes out on top. Ablations confirm both pieces of the recipe (contextual correction and contextual weighting) contribute, and Flux-OPD avoids the loss surges the authors observed when they tried to run OEL as a single continuous run.

The honest caveat is that the deltas are single-digit points on niche benchmarks, all teachers tested are from the Qwen family, and the paper doesn't quantify the extra compute cost of re-extracting contexts from 8 student rollouts every iteration versus vanilla OPD. Nor does it say how the method behaves outside the two open-ended tasks it targets. What is worth watching is who is on the author list: the Kling team ships one of the leading Chinese text-to-video systems, and a distillation recipe that specifically helps small prompt-rewriters align with a chosen video generator is exactly the sort of thing that lands in a production pipeline rather than staying in a paper.