NTU, CMU, Berkeley researchers release VBVR-Pro: a 300-task, verifiable-reward visual reasoning suite

Researchers from NTU, CMU, Berkeley, and other institutions have jointly released VBVR-Pro, a scalable and verifiable suite for native visual reasoning. The system defines 300 visual reasoning tasks, builds 1.25 million training data points, and implements verifiable scorers for 100 tasks that can serve directly as verifiable rewards in reinforcement learning. Experiments show that models trained with VBVR-Pro improve by up to 20 percentage points on unseen benchmarks, and the scorers achieve over 60% agreement with human judgment, outperforming tested VLMs.