Back to feed
News Story
机器之心
1 sources

NVIDIA, MIT, and Oxford Propose Physis-Lang to Enhance Video Physical Realism

NVIDIA, together with researchers from MIT and the University of Oxford, proposed Physis-Lang, which injects physical causes, laws, and outcomes into language descriptions to improve the physical realism of video models. On the Physics-IQ Verified leaderboard, Physis-Lang's Cosmos3-Super and Cosmos3-Nano versions ranked first and second with scores of 48.2 and 43.3, respectively, with Super surpassing the previous best by 5.5 percentage points. The team also built the PhysCapBench benchmark and a self-evolving caption prompt mechanism to continuously refine description guidelines based on physical evaluation feedback.

SynthePulse Insight · AI deep readingMembers

Writing Physics into Language: How Physis-Lang Lets Video Models 'Understand' Butter Melting

Version 1 · 1 source

NVIDIA, together with MIT and Oxford University, proposes Physis-Lang, which writes physical causes, laws, and outcomes into language descriptions and threads these descriptions through data filtering, model training, and video generation. On the Physics-IQ Verified leaderboard, its Super and Nano versions rank first and second with scores of 48.2 and 43.3, respectively. But what exactly does the language description change that allows the improvement to transfer to generated videos?

  • Physis-Lang was proposed by a research team from NVIDIA, MIT, and Oxford University. Its core idea is to write physical causes, laws, and outcomes into language descriptions and thread them through data filtering, model training, and video generation.
  • On the Physics-IQ Verified leaderboard, the Cosmos3-Super and Cosmos3-Nano versions of Physis-Lang rank first and second by average score with 48.2 and 43.3 points, respectively; Super exceeds the previous highest score of 42.7 by 5.5 percentage points.
  • After adding Physis-Lang, the 16B Cosmos3-Nano surpasses the original 64B Super model in average score.
Open section navigationWhat a Single Sentence 'Butter Melts When Heated' Omits

What a Single Sentence 'Butter Melts When Heated' Omits

Given the sentence 'butter melts when heated,' what should a video model generate? A piece of butter gradually shrinking, a ring of liquid spreading outward, or a yellow blob suddenly collapsing? This sentence points out the event but omits the intermediate process. How heat changes state, how gravity affects shape, and how solid shrinking and liquid spreading occur simultaneously all need to be filled in by the model itself.

Physis-Lang, proposed by a research team from NVIDIA, MIT, and Oxford University, starts from these often-omitted details: it writes physical causes, laws, and outcomes into language descriptions, and then threads these descriptions through data filtering, model training, and video generation. Taking butter as an example, a more complete description is that continuous heating causes the butter to gradually melt, the solid part collapses under gravity and continuously shrinks in volume, and the melted liquid forms a gradually expanding shallow pool around it. This passage provides continuous state changes. What the generative model needs to represent is no longer just the two concepts 'butter' and 'melting,' but also the effect of heating on the state of matter and the time-varying relationship between solid and liquid.

The Numbers on the Leaderboard: What 48.2 and 43.3 Mean

On the authoritative Physics-IQ Verified leaderboard, the Cosmos3-Super and Cosmos3-Nano versions of Physis-Lang rank first and second by average score with 48.2 and 43.3 points, respectively. Among them, Super exceeds the previous highest score of 42.7 shown in the figure by 5.5 percentage points; after adding Physis-Lang, the 16B Cosmos3-Nano also surpasses the original 64B Super model in average score.

These numbers come from the leaderboard page, and the article does not state third-party independent reproduction. Physis-Lang was proposed by a research team from NVIDIA, MIT, and Oxford University, and is a self-reported result by the research team. Therefore, the rankings and scores should be regarded as results under the team's specific evaluation conditions, not as independently verified conclusions.

Free for now

Read the full analysis

5 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

This article is based on Machine Heart's report on the Physis-Lang paper. Physis-Lang was proposed by a research team from NVIDIA, MIT, and Oxford University, and is a self-reported result by the research team. The Physics-IQ Verified leaderboard data comes from the leaderboard page, and the article does not state third-party independent reproduction. The atomic assertions in PhysCapBench were human-verified, but the verifiers and process details are not elaborated in the article. The comparison results on the four physics benchmarks come from the paper's evaluation, and the article does not give the complete experimental setup. Therefore, the rankings, scores, and improvement margins mentioned in the article should be regarded as reported results by the research team under specific evaluation conditions, not as independently verified conclusions.

Sources for this version

  1. 让自然语言成为世界模型的表征，英伟达Physis-Lang这样增强视频物理真实性机器之心

Primary report

机器之心

Primary source