UniEvo-VL: Multimodal Models Improve via On-policy Self-Distillation

Modern multimodal models can generate and understand content within a single system, allowing them to provide and learn from their own feedback. UniEvo-VL leverages this capacity by introducing a self-evolving framework where the model acts as both teacher and student, eliminating the need for a separate, larger external teacher.

Self-Distillation Mechanism

The framework uses the model's self-critiques as privileged information during training. The student model sees only the vanilla question, while the teacher model conditions on the privileged critique. Training minimizes the per-state divergence between their denoising diffusion distributions over the student's own sampling trajectories.

Performance Gains

Experiments conducted on the open-source Qwen-image-2512 demonstrated significant improvements in image generation capabilities:

• GenEval: Score increased from 0.747 to 0.808.
• GenEval2 Soft-TIFA: Score increased from 32.97 to 35.53.

Limitations and Observations

While the method improves overall generation, results indicate that self-improvement is not uniform across all tasks, particularly in mixed text-rendering outcomes. Additionally, tests with more powerful external critics like GPT5.6-Luna suggest that models with stronger judge capabilities can anticipate a higher ceiling for self-evolution.