VBVR-Pro: 300 visual reasoning tasks plus verifiable rewards for RL training

机器之心 · wechat · 2026-09-20

Researchers from NTU, CMU and Berkeley released VBVR-Pro, a scalable suite for native visual reasoning with papers, data, models and code all public:

300 visual reasoning tasks across perception, spatial, transformation, abstraction and knowledge, yielding 1.25M training samples rendered as both video and interleaved formats; trained models gain up to 20+ points on 7 unseen benchmarks.
Verifiable scorers for 100 tasks reach >60% agreement with humans, beating GPT-5.5 (0.54) and Gemini-3.1-Pro (0.52), nearing the 0.77 human ceiling.
Modality comparison over 30+ models: images suit end-state tasks, interleaved formats discrete states, video continuous dynamics; removing intermediate visual states hurts far more than removing text, revealing Chain-of-Step behavior.
RL: the scorers work as verifiable rewards — Wan2.2-TI2V-5B improves from 0.470 to 0.503 with SFT and 0.548 with RLVR, with gains on out-of-domain tasks.

More from Multimodal