VLMs Can Self-Improve as Visual Questioners Without External Supervision
June 10, 2026
A self-evolving framework lets a VLM act as both proposer and filter to generate harder, more visual-centric questions for its own training, avoiding collapse via diversity constraints. The model improves in both questioner and answerer modes without external data curation costs.
HOW THIS AFFECTS YOU
●
researcherThe proposer-filter self-training loop with diversity constraints is a reusable technique for bootstrapping question-generation capability in VLMs without labeled data.