4M-Param Video Model Emerges Physics Rules as a Controllable World Model
Author @mathemagic1an explores whether video models genuinely understand physics through experiments. By setting up a simple physics environment with collisions and wall bounces, they trained an approximately 400万参数 pixel Transformer to predict the next frame at a cost of just 10美元. The study shows that instead of merely performing pixel interpolation or memorizing the training set, the model internally forms representations of physical laws, which can be transformed into an interactive world model.
已确认
- 模型在中间层存在一个明显的“物理涌现区”，物理信息在此处变得可读。
- 将第5层激活按运动方向求平均并进行PCA后，会出现明显的圆环结构，且物理世界相邻的方向在模型内部表示中也保持邻近。
- 模型的激活可以通过线性映射直接读出X/Y坐标，并在更深层提前数步预测碰撞。
- 将特定碰撞场景留作未见样本，模型依然能较正确地模拟结果，这削弱了“只是背诵训练集”的解释。
- 借助Anthropic的Jacobian lens，成功从一个只看像素、没有动作标签的模型中提取出可写入的运动方向，并将其制作成可通过键盘控制的交互小游戏。
为什么重要
- 揭示黑盒机制：现代视频模型在处理对象恒存和遮挡时表现出强大的连续追踪能力。弄清楚这些物理信息如何在模型内部表征，有助于更好地理解和控制大型视频生成模型。
- 因果世界模型的潜力：实验表明，仅用原始像素训练Transformer，就有可能得到一个具备因果逻辑、甚至能像游戏一样被“玩起来”的世界模型，这为视频模型的实际应用与可控生成提供了新思路。
2026-07-29 ~ 2026-07-29 · 10 related posts
Primary sources
- Dissecting Video Models with Jacobian Lens: 4M Params Emerge Causal World Model — mathemagic1an · 2026-07-29
- Anthropic’s J-lens turns a raw-pixel video model into a playable world model — mathemagic1an · 2026-07-29
- Modern video models track object permanence and occlusion surprisingly well — mathemagic1an · 2026-07-29
- [source] A 4M-parameter pixel transformer learns a tiny physics game for about $10 — mathemagic1an · 2026-07-29
- The real question is whether video models memorize or learn physical law — mathemagic1an · 2026-07-29
- Held-out collision scenes still generalize in the video model — mathemagic1an · 2026-07-29
- Video-model activations linearly expose position and predict collisions early — mathemagic1an · 2026-07-29
- [source] Anthropic Jacobian lens pulls a writeable motion direction out of a pixel-only model — mathemagic1an · 2026-07-29
- Motion direction forms a ring in a video model’s activation space — mathemagic1an · 2026-07-29
- [source] Video models reveal a “Physics Emergence Zone” where motion direction becomes readable — mathemagic1an · 2026-07-29