2026-09-30 Daily Report: 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

Authors Fang Wu, Da Xing, Yanjie Huang, Junxi Wang, Ji Wang, Hejia Geng, Guancheng Wan, Bowen Zuo, Xiaomin Li, Shixiang Tang, Xinyu Xiang, Zehong Wang, Shiyi Du, Peng Xia, Shuangjia Zheng, Yining Hong, Li Erran Li, Jure Leskovec, Yejin Choi

Affiliations University of California, Riverside / Independent Researcher / Stanford University / The Johns Hopkins University / University of Oxford / University of Toronto / Carnegie Mellon University / The University of North Carolina at Chapel Hill / MatrAIx / University of Notre Dame

Categories Method / Self-Distillation / On-policy training recipe, Application / Multimodal Models / Image generation capability improvement, Evaluation / Model Improvement / Self-evolution framework evaluation

License CC BY 4.0

Abstract Overview

This paper presents UniEvo-VL, a self-evolution framework for multimodal models that uses their own visual critiques to improve later image generation. The method separates a student that generates from the original prompt and a teacher that sees a critique-augmented revised prompt, then matches their denoising predictions along the student’s own trajectories via on-policy self-distillation. The training pipeline can optionally filter revised prompts through a post-revision verification step before distillation. Experiments are conducted on compositional image generation and OCR-style text rendering using Qwen-Image-2512, with additional analysis of reflection at inference time and stronger external critics.

Novelty

The distinctive contribution is critique-conditioned on-policy self-distillation for multimodal self-improvement: the same multimodal system provides corrective feedback, turns it into privileged conditioning for a teacher, and distills that guidance into a student that only sees the original prompt. Unlike approaches that rely on corrected target images, scalar rewards, or a separately trained external teacher, UniEvo-VL learns from teacher predictions on the student’s own sampled states.

Results

Across the reported experiments, UniEvo-VL improves direct generation over the base model on GenEval and, with verification, also improves GenEval2 and OCR metrics. In Table 1, the verified Qwen configuration raises GenEval from 0.747 to 0.818, GenEval2 from 32.58 to 35.07, and OCR from 0.771 to 0.790, while a stronger external critic reaches the highest GenEval and GenEval2 scores overall. The paper also shows that reflection remains beneficial after training, and the combination of training plus reflection achieves the best native scores on all three benchmarks, although gains are uneven across tasks and prompt difficulty.

Key Points

UniEvo-VL trains a multimodal generator to internalize critique-derived corrections by distilling from a teacher conditioned on a revised prompt while the student keeps the original prompt.
A post-revision verification stage is used to filter synthesized revised prompts, and this verified setting yields stronger and more consistent improvements than the unverified self-critic configuration.
Empirically, improvements are strongest on harder compositional prompts, external critics can raise the improvement ceiling, and OCR/text-rendering gains are more mixed than compositional-generation gains.

References

arXiv: https://arxiv.org/abs/2609.38721v1
Fugu-MT: https://fugumt.com/fugumt/paper_check/2609.38721v1
Hugging Face Papers: https://huggingface.co/papers/2609.38721