MBZUAI IFM 开源物理世界模型 PAN 2，主打模拟式推理
MBZUAI 的 IFM 团队开源了物理世界模型 PAN 2，用 LLM 规划加 29B 扩散模型生成画面的两段式架构做行动前推演，去 LLM 消融直接掉到 8.7% 步骤正确率。做具身智能或智能体规划的人可以看看它的开源设计和数据。
MBZUAI 的 Institute of Foundation Models（IFM）发布了 PAN 2，一个面向物理智能的开源基础世界模型，核心思路是让智能体在行动前先在模拟中推演结果。架构分为两层：Conceptual 层由 LLM 负责推演会发生什么、何时发生以及镜头如何运动，并写成文本描述；Perceptual 层是一个从零训练的 29B 参数扩散模型，把这段文本转成可视画面。
消融数据说明了分工的重要性：去掉 LLM 后，多步任务只剩 8.7% 的步骤正确，没有任何一个计划能完整完成，说明动作顺序规划完全依赖 LLM。两层协同工作时，PAN 2 在 IFM 测试的模型中成绩最高，步骤正确率达 75.1%，完整完成计划的比例为 34.0%。由于每一步都必须成立计划才算成功，长序列任务的容错空间很小，这个结果也侧面反映了世界模型做长程规划的难度。
MBZUAI's Institute of Foundation Models (IFM) just released PAN 2, an open foundation world model for physical intelligence.
It's built towards simulative reasoning: letting an agent play out what happens before it acts.
It works with two kinds of reasoning, each with its own job:
> Conceptual: an LLM works out what happens, when, and how the camera moves, then writes it down as text. > Perceptual: a 29B diffusion model, trained from scratch, turns that text into what you see.
Take the LLM out, and multi-step tasks fall apart: 8.7% of steps right, no plan finished. Getting the order of actions right is the LLM's job.
With both, PAN 2 gets 75.1% of steps right and finishes 34.0% of plans in full, the highest among the models IFM tested. Every step has to hold for a plan to work.
More details in the technical blog:
https://t.co/l6xKPmKT9k
[1/1]
This report is categorized under AITOP's "Models" desk, focusing on Southeast Asia and the ASEAN-6 (Singapore, Malaysia, Indonesia, Thailand, Vietnam, Philippines) digital economy. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include PAN 2, MBZUAI, 世界模型, 扩散模型.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).
