Multimodal large language models (MLLMs) struggle with spatial, embodied, physical, and temporal reasoning. We hypothesize that these failures reflect a shared deficit in visual transition reasoning, and test whether this capability can serve as a shared training primitive that different models can learn from different supervision sources and reuse across different tasks, with a systematic training recipe. Existing benchmarks document these deficits separately but do not support controlled comparisons across scenes, actions, and reasoning operations. We therefore introduce WOVEN, a training source and benchmark for visual transition reasoning that organizes transition supervision by scene, action, and reasoning type, using diverse, realistic rollouts from video-pretrained generative models: 36,076 examples across 20 scene types, 5 action types, and 8 reasoning types.
Current MLLMs Show a Systematic Deficit
Using WOVEN, we first evaluate 38 frontier MLLMs (e.g., GPT-5.4 and Qwen3-VL-235B-A22B) and find a substantial and systematic deficit: even the strongest models fall far below humans, and the failures recur across model families and persist with scale. The best model, GPT-5.4, reaches 65.8% accuracy against a 92.3% human baseline. The gap also has a consistent structure across model families: for every model, temporal adjacency is harder than ordering, and counterfactual removal is harder than substitution. Accuracy drops more under geometric than under appearance changes, suggesting a systematic deficit in visual transition reasoning rather than an idiosyncrasy of any particular family or training recipe.
Visual Transition Reasoning Is Learnable
We then train MLLMs at multiple scales on WOVEN and find that they learn a shared capability that transfers broadly. Training subsets of only about 2,000 items each collectively improve 22 of 26 external benchmarks by up to 27.3 percentage points, and WOVEN data can replace 30-50% of a task's own training data with comparable accuracy. Training on the full WOVEN training set raises the in-distribution accuracy of Qwen2.5-VL-3B from 26.4% to 89.3%, with 88.2% on held-out scenes. The benefits extend to 12 of 26 external benchmarks across multiple domains. Controlled subsets containing approximately 2,000 examples each collectively expand this coverage to 22 of 26 benchmarks. The 44 benchmarks of static perception and general video QA do not improve under any training, so the gains are not due to an overall stronger model or to better multiple-choice answering.
Learned Transition Reasoning Transfers Broadly
Every controlled subset improves WOVEN accuracy, including on the action types it does not train on, meeting the requirement that different sources build the same capability. The more a subset improves WOVEN accuracy, the larger its average gain on the external benchmarks (r=0.86). Downstream gains result from the learned capability rather than from the individual subsets, as established by intervention on the training data: substituting up to 30-50% of task-specific items with WOVEN leaves accuracy comparable to training on these items alone, and above a control that simply deletes the replaced items.
Systematic Recipe for Training Visual World Modeling
Controlled comparisons yield a training recipe for visual world modeling, validated prospectively on held-out benchmarks. Select supervision by the reasoning operation it teaches rather than by the actions, scenes, or domains it shows. Once the operation is selected, choose supervision with larger changes to the visual state for robustness: actions that change more of the scene, such as object manipulation, yield larger robustness gains than actions that only move the camera. Supervise temporal and agent-driven tasks directly, as transfer does not cross the boundary between the temporal family and the other three reasoning families, nor the boundary between passive physical events and agent-driven actions. Keep dedicated supervision for capabilities beyond state transitions, as transition supervision improves diverse tasks that involve state transitions but does not necessarily generalize to capabilities such as static perception, which therefore need their own supervision.
From Dataset to Deployment
WOVEN establishes visual transition reasoning as a reusable foundation for improving diverse downstream capabilities, paving the way for systematic visual world-model training in MLLMs. The WOVEN project repository and dataset are available, and the constructed benchmarks and models provide a unified framework for advancing visual world modeling in MLLMs. The training recipe offers a controlled framework for selecting and combining supervision to maximize transfer across tasks, and the dataset spans the diversity of scenes, actions, and reasoning types needed for comprehensive evaluation and training.