IFM 开源世界模型 PAN 2，用模拟式推理让智能体先演练再行动
IFM 把世界模型做成了智能体的"预演器"：29B 的 LDM 负责画后果，LLM 负责想后果，选最优再动手。和纯文本推理的 agent 框架相比，它多了视频级的可视化验证。做机器人、游戏或决策型智能体的开发者可以直接下载权重和代码上手试。
Institute of Foundation Models（IFM）发布了世界模型 PAN 2，并提出"模拟式推理"（simulative reasoning）的思路：让智能体在执行前先模拟每个候选动作的后果，再选出最优方案。
视频示例中，模型对同一道浪的三个不同起乘时机分别推演，只有其中一个能成功追上浪，智能体因此选择该方案。技术上，PAN 2 采用 Generative Latent Prediction（GLP）：输入当前场景、动作和模拟时长，输出一段可供核查的视频。
推理分为两部分，由 LLM 承担概念推理，负责推演接下来会发生什么；再由一个 29B 参数的 LDM 承担感知推理，补全画面细节。
项目以 Apache 2.0 协议开源，权重、训练代码和推理代码全部公开，开发者可以直接运行并在其上构建。
This is a pretty cool idea:
Let your agent simulate each potential action before picking the best one.
The model is PAN 2, a world model from the Institute of Foundation Models (IFM). They call this "simulative reasoning".
The attached video shows an example with a surfer: same wave, three different moments to pop up. PAN 2 plays out all three, and only one of them catches the wave, so that’s the one he goes with.
The model uses Generative Latent Prediction (GLP), and here is how it works:
• In: the current scene plus an action, with how long to simulate
• Out: what happens next, as a video you can check
There are two reasoning parts:
1. Conceptual reasoning: The LLM is the backbone here. It reasons about what happens next.
2. Perceptual reasoning: A 29B LDM fills in what it looks like.
This is all open under Apache 2.0!
They provided the weights, training code, and inference code, so you can run it and build on it yourself.
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. 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 PAN 2, IFM, 世界模型, 开源模型.
Source attribution: curated from Ate-a-Pi (@svpino).
