UniEvo-VL: An On-policy Self-Distillation Training Recipe for Multimodal Model Self-improvement

UniEvo-VL. An On-policy Self-Distillation Training Recipe for Multimodal Model Self-improvement. By Fang Wu and eighteen others, from Stanford University, Johns Hopkins University, and eight other institutions. Published on the thirtieth of September. No code repository is linked to it.

Modern multimodal models put generation and understanding inside one unified system. One consequence is that such a model can look at what it has produced and comment on it, which means it can, in principle, give itself feedback and learn from it. The usual way to improve a model through distillation leans on a separate teacher, often a larger one. The question here is whether a unified model can serve as its own teacher, using its own constructive self-critique during test-time compute.

UniEvo-VL is a self-evolving framework built on that question. The model's self-critiques are treated as privileged information, and a single multimodal model is asked to play two roles under two different contexts. As the student, it sees only the vanilla question. As the teacher, it sees the same question together with the privileged critique. Same weights, different information.

Training then pulls the student toward the teacher. The objective minimizes the divergence, state by state, between the two roles' denoising diffusion distributions, and it does so over sampling trajectories that the student itself produced. That is what makes the recipe on-policy: the teacher is consulted on the paths the student actually takes, not on paths drawn from somewhere else. Nothing outside the model is needed at any point. The critique comes from the model, the teacher is the model, and the trajectories being corrected are the model's own. What differs between the two roles is only what each one is allowed to see.

The experiments report that UniEvo-VL improves the image generation of multimodal models while keeping them sensitive to additional reflection information. Built on the open-source Qwen-image-2512, it moves GenEval from zero point seven four seven to zero point eight zero eight, a gain the authors call significant. Keeping the sensitivity to reflection matters for the method's own logic: a model that stopped responding to critique would have nothing left to teach itself with.

On GenEval2, scored with Soft-TIFA, it moves from thirty-two point nine seven to thirty-five point five three. Trying stronger external critics, such as GPT5.6-Luna, suggests that models with strong judging ability can expect a higher ceiling for self-evolution. And the authors report a limit plainly: results on text rendering are mixed, so the self-improvement is not uniform across tasks.

The paper places itself in the current line of work on recursive self-improvement, with a practical aim: better results for people using multimodal models, without any external supervision or guidance. Its recipe comes down to one move. Let the model read its own critique, then teach the version of itself that did not read it. How far that goes appears to depend on how good a critic the model is, which is what the experiments with stronger external critics hint at, and the mixed text-rendering results are a reminder that the gains do not arrive evenly.