UniEvo-VL lets a multimodal model teach itself with its own critique, lifting GenEval from 0.747 to 0.808

The paper introduces UniEvo-VL, which turns visual critique into supervision: the understanding mode judges a generated image as accepted or needing revision and supplies corrective feedback, that feedback is synthesized into a revised prompt, and the revised prompt serves only as privileged conditioning for the teacher while the student still sees the original prompt. The authors note that prior self-enhancement, preference optimization, reflection tuning, or test-time policy optimization did not directly distill corrective conditioning into the original-prompt generation policy along its own sampling trajectories; UniEvo-VL uses two context roles of a single model instead of a separate, often larger, teacher. The method is formalized in Section 2, including the prompt-synthesis equation, teacher and student policy definitions, and the per-state divergence objective; implementation details, hyperparameters, and random seeds appear in Appendix A.

The training objective matches teacher and student local predictions at states along the student's own denoising trajectory, with gradients flowing only to the student and the teacher acting as a fixed full-distribution target, requiring neither corrected images as training targets nor scalar rewards. Unlike diffusion or flow optimization driven by reward gradients or image-level rewards, the targets here come from teacher predictions conditioned on corrective text, requiring neither gradients through the critic nor a differentiable image reward. The flow instantiation uses deterministic transition matching (velocity matching weighted by the squared sampler interval), and the paper states explicitly that this is not an exact stochastic-policy KL or JS divergence; only generator LoRA parameters are updated while base weights, text encoder, VAE, and feedback components stay fixed.

On Qwen-Image-2512, direct-generation performance rises from 0.747 to 0.808 on GenEval and from 32.97 to 35.53 on GenEval2 Soft-TIFA, and the configuration with post-revision verification exceeds the reported base on every listed metric across all three tasks. The paper separates gains retained in the model from the additional benefit of inference-time reflection and uses paired evaluation to show reflection still helps after training, indicating the acquired capability cannot be explained by inference-only prompt revision. Results come from paired evaluation on GenEval, GenEval2, and OCR, with Gemini-2.5-Flash used as an automated judge and HumanPref described as an automated visual-quality proxy rather than a human annotation study.

Critic capacity shapes the self-evolving ceiling: with GPT-5.6-Luna feedback the native GenEval score reaches 0.882 and GenEval2 reaches 35.53, whereas Qwen-VL feedback gives 32.37 on GenEval2, and text-rendering outcomes are mixed so gains are not uniform across tasks. The paper treats critic choice as a variable and reports category-level differences such as spatial position, counting, two-object composition, and color, noting that GenEval2 position and attribute categories improve only modestly. Category-level scores and Soft-TIFA GM/AM values are given in Table 3 and Appendix D.6; the authors also note these are historical subsets and specific configurations, with different configurations leading on individual metrics.

The work targets unified multimodal models that have both generation and understanding modes and that accumulate corrective experience from their own critiques at test time, applied to tasks with explicit textual specifications such as compositional image generation and visual text rendering. The validated instantiation is flow-based Qwen-Image-2512, where only generator LoRA parameters are updated and the critic and prompt-synthesis components stay fixed; the authors note an autoregressive extension would need a next-token distribution loss defined on a shared prefix and vocabulary. For practitioners this means the recipe lands most easily on models that already have a strong understanding branch and can produce executable corrective text, and that a stronger critic raises the expected self-evolving ceiling.

The reported results come from specific models and configurations, and different configurations lead on individual metrics, so a single number should not be read as a general conclusion. Gains are uneven across tasks: text-rendering outcomes are mixed, GenEval2 position and attribute categories improve only modestly, and training brings small regressions on initially easier prompts. Critique quality is itself a key variable, and the authors pose what makes a model's own feedback reliable enough to learn from as an open question; evaluation also relies on automated judges and an automated visual-quality proxy, and GenEval detector errors can penalize images that satisfy the prompt, all of which are worth keeping in view when interpreting the scores.