Signal VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning

Summary

Researchers introduce VBVR-Pro, a closed-loop testbed designed to make native visual reasoning — treating visual generation itself as the medium of reasoning — trainable, verifiable, and experimentally controllable. The suite turns visual reasoning into a controlled task space of 300 procedurally generated tasks, and models trained on it show transfer across seven external visual reasoning benchmarks. VBVR-Pro provides verifiable reward scorers grounded in deterministic, task-specific rules, and the authors identify recurring failure modes in the common practice of using vision-language models as judges. The suite also enables controlled modality studies across more than 30 image, video, and interleaved generators, finding video generation strongest for tasks needing persistent spatiotemporal state tracking. The paper was submitted to arXiv's cs.AI category on August 26, 2026. All data, models, scorers, and code are being released.