VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning
VBVR-Pro frames visual generation as the reasoning process, then scores it with deterministic task rules.
The suite contains 300 procedurally generated visual reasoning tasks built for training and controlled evaluation. The authors say models trained on it transfer to seven outside benchmarks, including RISE-Video, MME-CoF-Pro, and BabyVision. They argue its task-specific scorers avoid recurring failures they found in MLLM-as-judge setups and can serve as reinforcement-learning rewards. Their modality study compares more than 30 generators and finds video strongest for persistent spatiotemporal tracking, with interleaved generation as a cheaper alternative. ArXiv · AI/CL/LG's note
The suite contains 300 procedurally generated visual reasoning tasks built for training and controlled evaluation. The authors say models trained on it transfer to seven outside benchmarks, including RISE-Video, MME-CoF-Pro, and BabyVision. They argue its task-specific scorers avoid recurring failures they found in MLLM-as-judge setups and can serve as reinforcement-learning rewards. Their modality study compares more than 30 generators and finds video strongest for persistent spatiotemporal tracking, with interleaved generation as a cheaper alternative. ArXiv · AI/CL/LG's note
score 5