AI TrackerFrontier desk · AEST
Back to news
normalaitech

WOVEN: Weaving Visual World Modeling into Multimodal LLMs

arXiv cs.CL · 9 Oct 2026, 03:52 AEST

WOVEN introduces a training source and benchmark for visual transition reasoning in multimodal large language models (MLLMs). The study reveals significant deficits in existing models and demonstrates that training on WOVEN can enhance their performance across various benchmarks.

Key facts

Read aloud

WOVEN introduces a training source and benchmark for visual transition reasoning in multimodal large language models MLLMs. The study reveals significant deficits in existing models and demonstrates that training on WOVEN can enhance their performance across various benchmarks. Multimodal large language models MLLMs struggle with spatial, embodied, physical, and temporal reasoning. We introduce WOVEN, a training source and benchmark for visual transition reasoning that organizes transition supervision by scene, action, and reasoning type. We evaluate thirty eight frontier MLLMs and find a substantial and systematic deficit: even the strongest models fall far below humans. Training subsets of only about two thousand items each collectively improve twenty two of twenty six external benchmarks by up to twenty seven point three percentage points. Our work establishes visual transition reasoning as a reusable foundation for systematic visual world-model training in MLLMs.

Sources in this cluster