UniEvo-VL: An On-policy Self-Distillation Training Recipe for Multimodal Model Self-improvement
UniEvo-VL self-distills multimodal models from their own critiques, raising Qwen-Image GenEval scores.
UniEvo-VL is an on-policy self-distillation recipe in which one multimodal model is both teacher and student: the teacher conditions on a self-critique while the student sees only the question. Training minimizes divergence between their diffusion denoising distributions on the student's trajectories. On Qwen-Image-2512, GenEval rose from 0.747 to 0.808 and GenEval2 Soft-TIFA from 32.97 to 35.53. Stronger external critics such as GPT-5.6-Luna implied a higher ceiling, while text-rendering gains were uneven.
- One model acts as teacher and student with different contexts.
- Qwen-Image-2512 GenEval rose from 0.747 to 0.808.
- GenEval2 Soft-TIFA improved from 32.97 to 35.53.
- Text-rendering gains were uneven across tasks.
- AI models
- Qwen-Image-2512GPT-5.6-Luna
From the original article
1 min read at huggingface.coModern multimodal models bring generation and understanding into a single unified system, which enables them to provide and learn from their own feedback. Motivated by this unified capacity, we introduce UniEvo-VL, a self-evolving framework for multimodal models to learn from this constructive self-correction feedback during test-time compute.
By Fang Wu, Da Xing, Yanjie Huang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.