Keuntae Kim, Dongjun Kim, Yong Suk Choi · Dynamics-aware Video LLM
V-JEPA-LM: Video-Language Models That Read Dynamics Beyond Static, Text-Aligned Representations
The paper builds video-language models (1B to 8B, using Qwen3 and Llama) on the frozen V-JEPA world-model video encoder, with counterfactual supervision made from pixel-transformed public videos so the language model reads motion dynamics rather than retrieving captions. Trained in a single stage on 4.53M samples with 32 visual tokens per frame, the 4B model scores 72.1 across eleven video benchmarks, against 61.1 for the best open 4B and 64.8 for the best open 8B model. Counterfactual supervision adds 4.5 points on average at a matched budget. Code and models are on GitHub; the licence is not stated on the page.
Open the source ↗Similar Video & Animation releases
- MViS: Video Stitching from Multiple Moving Cameras
Ben-Gurion University of the Negev (Ifergane, Amoyal, Benishay, Freifeld) · Multi-camera video stitching · Week 47, 2026
- BirdDreams
NVIDIA (Junyoung Seo, Jaewoo Seo, Seungryong Kim, Laura Leal-Taixé, Xiuming Zhang, Seung Wook Kim) · Driving World Model · Week 47, 2026
- OverSeer
University of Oxford and Meta (Runjia Li, Andrea Vedaldi, Shikun Liu et al.) · Multi-agent world model · Week 45, 2026
- UniAction
Bowen Xue, Zizhang Li, Hong-Xing Yu, Jiajun Wu · Unified action world model · Week 45, 2026
- WorldCrafter
ARC Lab (Tencent IEG) & Peking University · Explorable World Model · Week 40, 2026
- Odyssey-3
Odyssey · World Model · Week 39, 2026
Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.