NVIDIA发布Physis-Lang：自进化物理语言提升Cosmos3视频生成
NVIDIA Researchers Introduce Physis-Lang: Self-Evolving Physical Language That Lifts Cosmos 3 Past Veo 3.1 on Physics Benchmarks
用纯文本提示工程实现视频物理一致性突破，无需修改架构即可超越Veo 3.1，对视频生成领域有重要参考价值。
Video world models can render convincing clips that still break physics. Butter spreads like paint. Balls pass through walls. A team from NVIDIA, MIT and the University of Oxford argues the fix can come from language itself, not from extra visual, latent or numerical signals.
视频世界模型可以生成逼真的片段，但仍会违背物理规律。黄油像油漆一样涂抹。球体穿过墙壁。来自 NVIDIA、麻省理工学院和牛津大学的一个团队认为，解决方案可以来自语言本身，而非额外的视觉、潜在或数值信号。
Their framework, Physis-Lang, treats physical language as a shared, optimizable representation. The same text drives data curation, model training and inference. On the public Physics-IQ Verified leaderboard snapshot dated September 29, 2026, Physis-Lang on Cosmos3-Super ranks first at 48.2 ± 1.4. The Cosmos3-Nano version ranks second at 43.3 ± 1.5.
他们的框架 Physis-Lang 将物理语言视为一种共享的、可优化的表示形式。相同的文本驱动数据策展、模型训练和推理。在截至 2026 年 9 月 29 日的公共 Physics-IQ Verified 排行榜快照中，运行在 Cosmos3-Super 上的 Physis-Lang 以 48.2 ± 1.4 的成绩排名第一。Cosmos3-Nano 版本以 43.3 ± 1.5 的成绩排名第二。
A video world model can make a convincing clip and still get the physics wrong.
视频世界模型可以制作出令人信服的片段，但仍然搞错物理原理。
Our researchers just released Physis-Lang, an open self-evolving framework that adds physics reasoning to video captions. The captions explain why and how a scene unfolds. We use them to fine-tune… pic.twitter.com/agIHuIB6N2
我们的研究人员刚刚发布了 Physis-Lang，这是一个开放的自进化框架，为视频字幕添加了物理推理。这些字幕解释了场景为何以及如何展开。我们利用它们进行微调…… pic.twitter.com/agIHuIB6N2
— NVIDIA AI (@NVIDIAAI) September 29, 2026
—— NVIDIA AI (@NVIDIAAI) 2026 年 9 月 29 日
What Problem Does Physis-Lang Solve?
Physis-Lang 解决了什么问题？
Conventional captions describe what happens, not why. ‘Butter melts as the temperature rises’ says nothing about heat transfer or gravity. Physis-Lang adds a physics_reasoning field to each base caption. It spells out entities, causes, interactions, governing principles, temporal evolution and effects.
传统的字幕描述发生了什么，而不是为什么发生。“随着温度升高，黄油融化”并未提及热传递或重力。Physis-Lang 在每个基础字幕中添加了一个 physics_reasoning 字段。它阐明了实体、原因、相互作用、支配原则、时间演变和效果。
The pipeline also writes a scene-specific physics_negative_prompt. This text describes likely implausible outcomes, such as a stone floating on water. It acts as negative conditioning at inference time.
该流水线还编写了特定于场景的 physics_negative_prompt。这段文字描述了可能不合理的结果，例如石头漂浮在水面上。它在推理时充当负向条件。
How Does the Self-Evolving Caption Loop Work?
自进化字幕循环是如何工作的？
The loop keeps the captioner frozen and evolves only its instruction. A GPT-5.5 captioner writes captions for a fixed 20-video development set with 273 human-verified assertions. Gemini-3.1-Pro acts as a physics-aware critic. An evolution agent reads the scores and claim-level failures, then rewrites the prompt.
该循环保持字幕生成器冻结，仅进化其指令。GPT-5.5 字幕生成器为固定的包含 273 个经人工验证断言的 20 个视频开发集编写字幕。Gemini-3.1-Pro 充当具备物理意识的评论家。进化代理读取评分和声明级别的失败情况，然后重写提示词。
The critic scores 2 dimensions:
评论家对两个维度进行评分：
- Precision: the caption is split into atomic claims, and each claim is checked against the video.
- Recall: each human-curated physical assertion must be explicitly stated or entailed by the caption.
- 精确度：字幕被拆分为原子声明，每个声明都与视频进行核对。
- 召回率：每个由人工策划的物理声明必须在字幕中明确陈述或由字幕蕴含。
Every revised prompt is validated on PhysCapBench, a new benchmark of 246 videos and 3,794 human-verified assertions. Caption F1 rose from 78.64 at iteration 1 to 87.82 at iteration 9. The path was not smooth. Iteration 2 made captions overly cautious and dropped F1 to 76.28. Iteration 9 required every visible causal step and raised frame sampling from 2 fps to 4 fps.
每次修订后的提示词都在 PhysCapBench 上进行验证，这是一个包含 246 个视频和 3,794 个经人工验证断言的新基准测试。字幕 F1 分数从第 1 次迭代的 78.64 上升到第 9 次迭代的 87.82。过程并非一帆风顺。第 2 次迭代使字幕过于谨慎，F1 分数降至 76.28。第 9 次迭代要求涵盖每一个可见的因果步骤，并将帧采样率从 2 fps 提高到 4 fps。
How Does Language-Guided Data Curation Work?
语言引导的数据策展是如何工作的？
A GPT-5.5 diagnosis agent maps generated-video failures to physics categories like rigid-body motion, collision and fluid dynamics. That deficiency profile is matched against physics tags on a large video gallery. Retrieval targets physical content, not visual appearance.
GPT-5.5 诊断代理将生成视频中的失败案例映射到刚体运动、碰撞和流体动力学等物理类别。该缺陷特征档案与大型视频库中的物理标签进行匹配。检索目标是物理内容，而非视觉外观。
The final training set holds 183K videos: 71K filtered from WISA-80K plus 112K retrieved clips. Retrieval alone added 3.01 points on average across 3 benchmarks. On VideoPhy-2, chemical and thermal processes each gained 8.00 points.
最终训练集包含 183K 个视频：其中 71K 个是从 WISA-80K 中筛选出来的，另外 112K 个是检索到的片段。仅通过检索，在 3 个基准测试上的平均分就提升了 3.01 分。在 VideoPhy-2 上，化学过程和热过程各自提升了 8.00 分。
How Does Physis-Lang Compare With Veo 3.1?
Physis-Lang 与 Veo 3.1 相比如何？
Fine-tuning uses LoRA on attention projections, with no architecture or objective change. Physis-Lang on Cosmos3-Nano versus Google's Veo 3.1:
微调仅在注意力投影上使用 LoRA，不改变架构或目标函数。Physis-Lang 在 Cosmos3-Nano 上与 Google 的 Veo 3.1 对比：
- PhyGenBench: 71.04 vs 65.63
- Physics-IQ Verified: 43.41 vs 34.99
- PhyGround: 69.90 vs 69.24
- VideoPhy-2: 68.02 vs 68.87 on the full set, 62.36 vs 58.43 on the Hard split
- PhyGenBench：71.04 vs 65.63
- Physics-IQ Verified：43.41 vs 34.99
- PhyGround：69.90 vs 69.24
- VideoPhy-2：全量数据集上为 68.02 vs 68.87，Hard 分割集上为 62.36 vs 58.43
Gains hold across backbones: +7.05 on Wan2.1-14B, +3.24 on Cosmos3-Edge-4B, +6.22 on Cosmos3-Nano-16B and +5.02 on Cosmos3-Super-64B. General quality held steady on VBench-I2V, where Cosmos3-Nano moved from 88.32 to 88.69.
增益在不同骨干网络上均保持一致：Wan2.1-14B 上提升 +7.05，Cosmos3-Edge-4B 上提升 +3.24，Cosmos3-Nano-16B 上提升 +6.22，Cosmos3-Super-64B 上提升 +5.02。在 VBench-I2V 上，整体质量保持稳定，其中 Cosmos3-Nano 从 88.32 提升至 88.69。
Prompting alone also helps. Physics reasoning plus negative prompts lifted a frozen Cosmos3-Nano on PhyGenBench from 61.67 to 67.29.
仅使用提示词也能带来帮助。结合物理推理和负向提示词，使冻结状态的 Cosmos3-Nano 在 PhyGenBench 上的得分从 61.67 提升至 67.29。
Can It Run Without Commercial APIs?
它能否在不依赖商业 API 的情况下运行？
The research team distilled the GPT pipeline into 2 Qwen3-VL-4B-Instruct models: PhysThinker-C for captioning and PhysThinker-U for prompt upsampling. On Wan2.1-14B, the commercial pipeline gave +7.05 at about $24.12K in API cost. Swapping in PhysThinker-C kept +6.76 at about $0.12K. A fully local setup cost $0 and still added +4.76.
研究团队将 GPT 流程蒸馏为两个 Qwen3-VL-4B-Instruct 模型：用于字幕生成的 PhysThinker-C 和用于提示词上采样的 PhysThinker-U。在 Wan2.1-14B 上，商业管道带来了 +7.05 的提升，API 成本约为 $24.12K。替换为 PhysThinker-C 后，保持了 +6.76 的提升，成本降至约 $0.12K。完全本地化的设置成本为 $0，但仍能带来 +4.76 的提升。
Physis-Lang vs Closest Competitors
Physis-Lang 与最接近的竞争者对比
| Feature | Physis-Lang | PhiZero | PhyGDPO | Self-Refinement |
|---|---|---|---|---|
| Developer | NVIDIA, MIT, Oxford | CASIA (NLPR) | Meta (ECCV 2026) | Liu et al. |
| Core idea | Self-evolving natural-language physics captions and negative prompts | Learned discrete "physical language", reason-then-render | Groupwise DPO with VLM physics rewards | Multimodal chain-of-thought prompt refinement from VLM feedback |
| Where physics enters | Data curation, training captions and inference prompts | Qwen3-VL-4B reasoner feeding a diffusion decoder | Preference training on PhyVidGen-135K | Inference prompts only |
| Training needed | LoRA SFT (prompt-only mode also helps) | Yes | Yes (DPO) | No, training-free |
| Physics-IQ Verified | 43.41 | 40.91 | n/r | 27.20 |
| PhyGenBench | 71.04 | n/r | 48.96 | 49.17 |
| VideoPhy-2 (All / Hard) | 68.02 / 62.36 | n/r | 59.56 / 44.94 | 47.88 / 28.09 |
| PhyGround | 69.90 | 57.85 | n/r | 58.22 |
| Code / weights public | Paper only | "Coming soon" | "Released soon" | Paper |
| 特性 | Physis-Lang | PhiZero | PhyGDPO | Self-Refinement |
|---|---|---|---|---|
| 开发者 | NVIDIA、MIT、牛津大学 | 中科院自动化所（NLPR） | Meta (ECCV 2026) | Liu 等人 |
| 核心理念 | 自我进化的自然语言物理描述与负面提示 | 学习离散的“物理语言”，先推理后渲染 | 基于 VLM 物理奖励的组级 DPO | 从 VLM 反馈中进行多模态思维链提示优化 |
| 物理介入环节 | 数据策展、训练描述与推理提示 | Qwen3-VL-4B 推理器驱动扩散解码器 | 在 PhyVidGen-135K 上进行偏好训练 | 仅用于推理提示 |
| 是否需要训练 | LoRA SFT（仅提示模式也有帮助） | 是 | 是（DPO） | 否，免训练 |
| Physics-IQ 验证得分 | 43.41 | 40.91 | n/r | 27.20 |
| PhyGenBench | 71.04 | n/r | 48.96 | 49.17 |
| VideoPhy-2 (全部/困难) | 68.02 / 62.36 | n/r | 59.56 / 44.94 | 47.88 / 28.09 |
| PhyGround | 69.90 | 57.85 | n/r | 58.22 |
| 代码/权重公开情况 | 仅有论文 | “即将推出” | “即将发布” | 仅有论文 |
Scores are from the Physis-Lang paper, Tables 1 to 4, run under one protocol per benchmark (PhyGenBench and VideoPhy-2 use a GPT-5.5 evaluator). Physis-Lang numbers use the Cosmos3-Nano backbone. n/r = not reported in that comparison. Release status checked September 30, 2026.
分数来源于 Physis-Lang 论文的第 1 至第 4 表，每个基准测试均采用单一协议运行（PhyGenBench 和 VideoPhy-2 使用 GPT-5.5 评估器）。Physis-Lang 的数据基于 Cosmos3-Nano 骨干网络。n/r = 在该对比中未报告。发布日期状态截至 2026 年 9 月 30 日核查。
Key Takeaways
关键要点
- Physis-Lang evolves physics captions with a critic-guided agent while the captioner stays frozen.
- PhysCapBench scores captions on 3,794 human-verified cause, law and effect assertions.
- Cosmos3-Nano with Physis-Lang beats Veo 3.1 on 3 of 4 benchmarks.
- Physics prompts alone lift a frozen model by 5.62 points on PhyGenBench.
- No code or weights are public yet; only the paper is released.
- Physis-Lang 通过批评家引导的智能体进化物理描述，而描述生成器保持冻结状态。
- PhysCapBench 对 3,794 个经人类验证的因果、定律及效果断言进行评分。
- 搭载 Physis-Lang 的 Cosmos3-Nano 在 4 项基准测试中的 3 项上超越了 Veo 3.1。
- 仅凭物理提示即可使冻结模型在 PhyGenBench 上提升 5.62 分。
- 目前尚无公开的代码或权重；仅发布了论文。
Check out the Paper, Project Page and GitHub Repo. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
查看论文、项目页面和 GitHub 仓库。所有荣誉归于该项目的研究者。此外，欢迎在 Twitter 上关注我们，并别忘了加入我们拥有超过 15 万成员的 ML SubReddit 以及订阅我们的新闻通讯。等等！你在 Telegram 上吗？现在你也可以加入我们的 Telegram 群组。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力