WOVEN: Weaving Visual World Modeling into Multimodal LLMs
The paper argues that visual transition reasoning is the missing shared skill behind several MLLM reasoning failures.
WOVEN adds 36,076 examples spanning scenes, actions, and reasoning types, built from realistic rollouts by video-pretrained generative models. The authors test 38 frontier multimodal models and report that even the strongest remain well below humans on these transition tasks. Training on WOVEN transfers to 22 of 26 external benchmarks, with gains reported up to 27.3 points. Their recipe favors supervision by reasoning operation, with larger visual-state changes helping robustness. HF Daily Papers' note
WOVEN adds 36,076 examples spanning scenes, actions, and reasoning types, built from realistic rollouts by video-pretrained generative models. The authors test 38 frontier multimodal models and report that even the strongest remain well below humans on these transition tasks. Training on WOVEN transfers to 22 of 26 external benchmarks, with gains reported up to 27.3 points. Their recipe favors supervision by reasoning operation, with larger visual-state changes helping robustness. HF Daily Papers' note
score 5