This paper introduces VBVR-Pro, a testbed that trains and measures “native visual reasoning.” That means the system uses images and videos themselves as the medium for thinking, not only as inputs to analyze or outputs to show. The authors built a closed-loop environment so generation, evaluation, and learning happen together.

VBVR-Pro turns visual reasoning into a controlled set of 300 procedurally generated tasks. It also provides verifiable reward scorers. These scorers use deterministic, task-specific rules to evaluate whether a generated image or video meets the task goal, instead of relying on another large model to judge the result.

The team studied how different judging methods work. They ran a systematic study of leading multimodal large language models (MLLMs) used as judges and found recurring failure modes in the common practice of using vision-language models as judges. In contrast, the rule-based scorers in VBVR-Pro match human judgments more closely and serve as reliable reward signals for large-scale multi-task reinforcement learning (RL).

The authors used VBVR-Pro to compare more than 30 generators across images, videos, and interleaved generation (switching between image and video). They report that video generation is strongest for tasks that need persistent spatiotemporal state tracking—keeping track of things that change over time—while interleaved generation is a more compute-efficient alternative. Ablations and probing experiments in the suite point to the importance of “vision-native trajectories,” which are sequences of visual states that seem crucial for solving these reasoning tasks.

VBVR-Pro also shows practical transfer: models trained on the suite showed strong transfer to seven external visual reasoning benchmarks, including RISE-Video, MME-CoF-Pro, and BabyVision. The authors release their data, models, scorers, and code. At the same time, they highlight limits in the current judging approach and emphasize that their deterministic, task-specific scorers are central to reliable training and evaluation within this controlled testbed.