ResearchSeptember 30, 2026
27
š¶ļø
NVIDIA Teaches Video Models Physics via Language
NVIDIA researchers present Physis-Lang, a self-evolving physical language that lifts Cosmos 3 past Veo 3.1 on physics benchmarks.
#NVIDIA#World Models#Video Generation#Physics AI#Research
š„ What happened
NVIDIA, MIT and Oxford dropped Physis-Lang, a framework that injects physics reasoning into video captions. On the Physics-IQ leaderboard (Sept 29, 2026), Cosmos3-Super takes first place at 48.2.
š” Why it matters
Language ā not extra visual signals ā drives data curation, training and inference. Cosmos3-Nano beats Veo 3.1 on 3 of 4 benchmarks, and the fully local setup costs $0 versus $24K in API fees.
ā” Our take
Proof that caption-level prompt engineering beats expensive architecture tricks. But no weights, no code ā just a paper. Until then, it's a pretty demo.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions ā when in doubt, read the linked original source.
