MMGR: Multi-Modal Generative Reasoning Benchmark and Evaluation
Evaluating state-of-the-art video generators, image generators, and LLM/VLM baselines reveals a sharp gap between visual quality and reasoning correctness: video models perform best on Physical Commonsense, but remain weak on symbolic tasks such as Sudoku, ARC, and Math, and brittle in cross-view embodied navigation.
Ze-Fan Cai, Haoyi Qiu, Tian-Yi Ma et al.
· 9 citations
· ⚡1