Program-Verified Self-Evolution for Vision-Language Models

Program-Verified Self-Evolution for Vision-Language Models. By Ahmed Heakl and three others, from L G A I Research, MBZUAI, and two other institutions. Published on the twenty-seventh of September. Its code repository is public and real: about twenty-five source files.

A self-evolving vision-language model trains on questions it writes for itself about unlabeled images. Those questions have no gold answers, so earlier methods label them in one of two ways: a majority vote over several sampled answers, or a verdict from a model acting as judge. The authors checked those labels with human evaluation, and found that twenty-four percent of the majority-vote labels, and eighteen percent of the judge labels produced during self-evolution, are wrong. A model learning from its own answers is learning a sizeable share of mistakes.

Verifiable QA Generation for Self-Evolving Models, or VQS, changes how answers get decided. Rather than vote on an answer, the model first turns each image into a structured record: a scene graph, for example, or a table read off a chart, or a graph of a diagram. From that record, fixed programs write the question and compute its answer. The work of reaching an answer moves out of the model and into code.

The model keeps one job as a visual checker. It confirms, one short claim at a time, the individual facts that the program read from the record. Those claim-level checks also decide which parsed records become training targets for the parser, so the model's ability to turn images into records improves too, still without any labels. The two roles feed each other. Better parsing gives the programs better records to work from, and the checks decide what the parser learns next.

Human raters judge ninety-four percent of the answers VQS produces to be correct, compared with seventy-six percent under majority voting. Across ten benchmarks, VQS improves Qwen3-VL by as much as three point one eight points, at each of the two, four and eight billion parameter scales, and it beats the strongest self-evolving baseline at every one of those sizes. The gains keep climbing over three rounds of training, reaching three point eight four points at the smallest scale. The code has been released.

What changes is the kind of question the model is asked to answer about its own work. Votes and judges ask it to settle a whole answer at once, and they get a large share wrong. VQS asks for something narrower: confirm one short claim at a time, and leave the computing of the answer to a fixed program. By the human raters' count, that narrower question yields correct labels far more often.