UniEvo-VL: An On-policy Self-Distillation Training Recipe for Multimodal Model Self-improvement
An image model critiques its own pictures, then trains itself to get them right the first time.
0.808GenEval score (0–1)
GenEval score after self-training
No external teacher or reward model needed for the gain
How it works
Critiques help only once
A unified model can spot flaws in its own images, but fixing them needs an extra reflection pass every time.
Same model, two roles
The teacher copy sees a critique-revised prompt; the student sees the original. The student matches the teacher's denoising steps.
Better first drafts
Direct generation improves on GenEval, GenEval2 and OCR. Reflection still adds gains on top.
What they found
Which comes out on top on GenEval native score?
Self-training over updates
try itDrag the number of optimizer updates and watch the direct-generation score.
- GenEval score
- medium
- here
- base model
Terms
6 termsPrivileged information The critique-revised prompt only the teacher sees during training.
A cook tastes a dish, writes notes, cooks it again following the notes, then practices until the first attempt matches the note-guided version.
Limits
- Text-rendering gains are mixed; unverified setup falls below base on OCR.
- Easy prompts regress by 1–4 points on a 0–100 scale.
- Biggest gains need a stronger external critic, not pure self-feedback.