WOVEN: Treating World Modeling as a Measurable Standalone Capability
What happened
A team from Northwestern, CMU, UNC Chapel Hill, and Amazon, in the paper "WOVEN: Weaving Visual World Modeling into Multimodal LLMs," defines world modeling as estimating the state transition p(s′|s,a) and builds WOVEN, a training and evaluation base containing 36K questions. The team tested 38 frontier models; the strongest performer, GPT-5.4, answered 65.8% correctly, versus 92.3% for humans. The team reports that each controlled subset, using only 2,000 data points, improves 22 benchmarks across domains; post-training on about 7 hours of generated video yields a gain of +6.7 on 3DSRBench, higher than Orca-4B's +4.0 from mid-training on 125,000 hours of real video. The team also notes that temporal tasks, agent-driven tasks, and static perception are areas where transition supervision is difficult to help directly or offers limited help.
AI 根据报道生成 · 4 days ago更新
Reporting timeline
Follow the reporting to explore different perspectives.
- 机器之心 · 微信公众号FeaturedNorthwestern, CMU and others propose WOVEN: cross-domain gains on 22 benchmarks with 2k samples, plus a training recipe
A team from Northwestern, CMU, UNC Chapel Hill and Amazon, in the paper "WOVEN: Weaving Visual World Modeling into Multimodal LLMs," defines world modeling as estimating the state transition p(s′|s,a) and builds WOVEN, a training and evaluation base of 36K questions.
Activity for this event
- Current
- 3
- Peak
- 6Nov 7 01:00
- Past 24 hours
- -49%
The trend compares participants observed throughout the same period, so its range may be smaller than the current interest count. Move or tap to inspect each hour; use the arrow keys on a keyboard.