WOVEN: Weaving Visual World Modeling into Multimodal LLMs
WOVEN trains multimodal LLMs on visual transitions, lifting 22 of 26 external benchmarks by up to 27.3 points.
Researchers introduce WOVEN, a visual transition-reasoning source and benchmark with 36,076 examples across 20 scene types, five action types, and eight reasoning types, drawn from video-pretrained generative rollouts. Evaluating 38 frontier multimodal LLMs, including GPT-5.4 and Qwen3-VL-235B-A22B, shows a large gap versus humans that persists across families and scale. Training on subsets of about 2,000 items improved 22 of 26 external benchmarks by up to 27.3 percentage points, and WOVEN data could replace 30-50% of a task's own training data. The recipe selects supervision by the reasoning operation taught and favors larger changes to the visual state.
- 36,076 examples span 20 scene types, five actions, and eight reasoning types.
- Thirty-eight frontier MLLMs, including GPT-5.4, remain far below humans.
- About 2,000-item subsets improved 22 of 26 benchmarks by up to 27.3 points.
- WOVEN data can replace 30-50% of a task's own training set.
- Select supervision by reasoning operation and prefer larger visual-state changes.
Full article273 words · extracted from arxiv.org · click to collapse
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, one 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. 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. 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. Controlled comparisons further 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, and prefer larger changes to the visual state for robustness. Our work establishes visual transition reasoning as a reusable foundation for systematic visual world-model training in MLLMs.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2610.12417