A multimodal model that tries to improve itself usually gets very little to learn from: a single coarse score for a whole answer, right or wrong. Distillation gives a much richer signal, correcting the model token by token rather than answer by answer, but in the visual domain it has normally needed outside help to work — human annotations, external tools, or a stronger model to imitate.

This paper proposes doing it with no outside help at all. The method, CVPD, hunts for the model’s own blind spots: image regions where zooming in changes and sharpens its answer, yet deleting the same region barely changes what it says about the full picture. The authors read that combination as information the model can already perceive but fails to use consistently. A three-part counterfactual test finds those regions from the model’s own responses, and they become dense training signal — the model teaches itself from what it was already overlooking.

Tested on Qwen3-VL-8B-Instruct, they report beating six self-evolving baselines across twelve benchmarks with no regression on any of them, including methods that lean on GPT-4o for supervision. The specific gains cited are +3.60 on OCRBench, +3.38 on MMStar fine-grained perception, and +3.08 on MMStar logical reasoning.

Worth reading with the usual care: this summary is drawn from the paper’s abstract, so the numbers above are the authors’ own reported results on a single 8B model. Whether the blind-spot criterion generalises to other model families, or holds up under independent replication, is not something the abstract establishes.

Written by the Hevolve AI agent from this paper's abstract, and reviewed by a person before publication. The abstract is quoted below so you can check it against the source.