Papers

BAAI 发布 PhysCapBench:首个评估视频字幕物理保真度的跨域基准

Why we picked it

做视频生成或具身智能数据的人应该马上看这个:智源联合北大清华发了个基准,专门查视频字幕里有没有写错重力、碰撞这些物理细节。代码、数据集、榜单全开源,能直接拿来给自己的字幕数据做体检,这在之前的评测体系里是空白。

BAAI(北京智源人工智能研究院)联合北京大学、清华大学、北京航空航天大学和 Physis 发布 PhysCapBench,这是首个跨领域基准,用于衡量视频字幕在多大程度上保留真实世界的物理信息,服务对象是 World Models 和 Embodied AI 的数据筛选。核心论点是:字幕语言流畅不等于物理接地,一段文采斐然的视频描述可能在重力、碰撞、运动轨迹等关键物理事实上有错。

Body

基准配套了完整开源资产:评测代码、榜单和数据集均已在 Hugging Face 等平台公开,同时附有 ArXiv 论文说明评测方法。对做视频生成模型训练或具身智能数据管道的团队而言,可以用它量化检查现有字幕数据的物理准确性,而不是只靠人工抽检。

对行业的影响在于数据标准:World Models 的进展越来越依赖高质量视频-文本对,PhysCapBench 把'物理一致性'变成了可排名、可复现的评测维度,类似当年视频理解领域 SOTA 榜单的做法。

Original · 智源研究院 BAAI

Fluent captions ≠ Physical grounding. 🚨

Today, BAAI introduce PhysCapBench — the first cross-domain benchmark measuring how well video captions preserve real-world physics for World Models & Embodied AI.

Everything is open-sourced to accelerate physically grounded AI data curation:

🔗 Project & Leaderboard: https://t.co/2jfhrU5Lfk 💻 Evaluation Code: https://t.co/Dt2cVX2qqU 🤗 Benchmark Dataset: https://t.co/1oEDfC1utK 📄 ArXiv Paper: https://t.co/yFFZyPBYna

Joint work by @BAAIBeijing, @PKU1898 , @Tsinghua_Uni , Beihang University, and Physis.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include PhysCapBench, BAAI, World Models, embodied AI.

Source attribution: curated from 智源研究院 BAAI (@BAAIBeijing).