Self-Questioning Vision-Language Models: Reinforcement Learning for Compositional Visual Reasoning

Researchers have proposed a self-questioning framework that trains vision-language models to independently break complex visual questions into smaller sub-questions, improving compositional reasoning without relying on expensive human-written explanations [1]. The work, posted on the arXiv preprint server on June 14, 2026, addresses a persistent weakness in Vision-Language Models (VLMs) — AI systems that process both images and text [1][2]. These models often struggle with compositional visual reasoning questions that require chaining multiple steps, such as identifying objects, counting them, and comparing the results [2]. Existing methods improve this reasoning by training models on human-written step-by-step explanations, but creating those annotations is expensive and difficult to scale [1][2]. The new framework, developed by Saraswathy Amjith, instead trains a VLM to generate its own intermediate sub-questions and answer each one before producing a final response [1]. The model is never shown examples of how to decompose questions; it discovers this behavior on its own, guided by a reward signal that scores whether the output contains sub-questions and whether the final answer is correct [1][2]. The training uses a reinforcement learning algorithm called Group Relative Policy Optimization, or GRPO [1][2]. The researchers applied the framework to a 3-billion-parameter model and tested it on both synthetic scenes of geometric shapes from the CLEVR dataset and real-world photographs from the A-OKVQA benchmark [1][2]. On A-OKVQA, the self-questioning model reached 52.2% accuracy, compared with 51.6% for standard reinforcement learning and 46.8% for the untrained model [1][2]. The paper describes this as the first self-questioning VLM, achieved by rewarding not only the final answer — as standard reinforcement learning does — but also the generation of intermediate sub-questions, enabling the model to discover compositional decomposition strategies [2]. The preprint appeared on arXiv, an open-access repository that hosts scientific papers across mathematics, physics, computer science, and related fields and that, as of late 2024, receives about 24,000 submissions per month [6]. The findings suggest that teaching AI systems to ask themselves intermediate questions is a promising strategy for complex visual reasoning, particularly when the difficulty of a question warrants explicit step-by-step decomposition [2].