VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning
VBVR-Pro makes visual generation itself trainable as a reasoning process, with rule-based rewards instead of VLM judges.
The paper introduces a closed-loop testbed built around 300 procedurally generated native visual reasoning tasks. Its scorers use deterministic, task-specific rules and are presented as better aligned with human judgment than common VLM-as-judge setups. Models trained on the suite reportedly transfer to seven outside visual reasoning benchmarks. The authors also compare more than 30 image, video, and interleaved generators, finding video strongest for persistent spatiotemporal tracking and interleaved generation a cheaper alternative. HF Daily Papers' note
The paper introduces a closed-loop testbed built around 300 procedurally generated native visual reasoning tasks. Its scorers use deterministic, task-specific rules and are presented as better aligned with human judgment than common VLM-as-judge setups. Models trained on the suite reportedly transfer to seven outside visual reasoning benchmarks. The authors also compare more than 30 image, video, and interleaved generators, finding video strongest for persistent spatiotemporal tracking and interleaved generation a cheaper alternative. HF Daily Papers' note
score 4