Self-Evolving Visual Questioner

A team of researchers has proposed a self-evolving framework that allows vision-language models to improve as visual questioners without any external supervision, according to a paper submitted to arXiv on 11 Jun 2026 [1]. Vision-language models, or VLMs, are typically trained as passive answerers, while their ability to actively ask diverse, non-trivial, visual-centric and grounded questions remains underexplored [1]. Existing visual questioners' performance is bottlenecked by the availability of high-quality training data or the cost of curating them [1]. The new framework uses a VLM itself as both a proposer and a filter to produce harder, more informative, and visual-centric questions, while maintaining their exploration diversity to avoid training collapse [1]. These questions are then used to train the VLM in both questioner and answerer modes [1]. To evaluate the questioner, the researchers introduce an agentic protocol that assesses questions along perception, reasoning, and diversity dimensions [1]. Experiments across various backbone VLMs show that the method substantially enhances the quality and substantially expands the difficulty boundary of autonomous question generation [1]. Under the same budget, the self-supervision is more effective than training on the static source data [1]. Moreover, the self-evolving questioner remains a competitive or even better answerer [1]. The paper was posted on arXiv, an open-access repository of electronic preprints that, as of November 2024, receives about 24,000 articles per month [9]. The work appears within the Computer Vision and Pattern Recognition section, a field where large language models with many parameters are often trained with self-supervised learning on vast amounts of text [11]. The abstract page for the paper includes a link to Hugging Face, a platform commonly used to share machine-learning models and datasets [1]. The research was disseminated through arXivLabs, a framework that allows collaborators to develop and share new arXiv features directly on the website, with partners that have accepted values of openness, community, excellence, and user data privacy [7].