Causal Physics Steering in Video World Models via Concept Activation Vectors

• URL: http://arxiv.org/abs/2605.24322v1
• Date: Sat, 23 May 2026 01:02:11 GMT
• Title: Causal Physics Steering in Video World Models via Concept Activation Vectors
• Abstract summary: Video models learn representations of physical dynamics, but controlling their physical expectations at inference time remains an open problem.<n>Recent interpretability work identified a Physics Emergence Zone (PEZ) where physical plausibility is represented separately from other visual features.<n>We present physics steering, a training-free method that uses the weight vector of a linear probe at a PEZ layer as a Concept Activation Vector (CAV) and injects it into hidden states during inference.
• Score: 4.111899441919164
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Video world models learn representations of physical dynamics, but controlling their physical expectations at inference time remains an open problem. Recent interpretability work identified a Physics Emergence Zone (PEZ), a group of middle transformer layers in VideoMAE where physical plausibility is represented separately from other visual features. However, it remained unclear whether this structure could be used to directly control the model's physics reasoning. We present physics steering, a training-free method that uses the weight vector of a linear probe at a PEZ layer as a Concept Activation Vector (CAV) and injects it into hidden states during inference. This shifts the model's physical expectations without changing any model weights. On the IntPhys benchmark, this intervention reliably shifts the model's plausibility judgment in either direction, depending on the steering sign. The effect appears only when the intervention is applied within the Physics Emergence Zone, suggesting that the relevant physics representation is localized there. We further find that physics is encoded separately from motion direction, and that different intuitive physics principles occupy distinct directions within this representation space. Together, these results show that physical reasoning in VideoMAE is not only readable, but also directly steerable.

Related papers

• PhysiFormer: Learning to Simulate Mechanics in World Space [56.16644631495]
  We present PhysiFormer, a diffusion transformer for physically-plausible 3D object motion.<n>It represents objects as 3D meshes expressed in world coordinates.<n>It substantially outperforms autoregressive baselines in trajectory accuracy, rigidity preservation, and momentum-based physical consistency.
  arXiv  Detail & Related papers  (2026-06-25T17:59:00Z)
• Neural Voxel Dynamics: Learning Implicit 3D Physics via Volumetric Feature Advection [0.14323566945483496]
  We present a self-dimensional framework for learning implicit 3D physical dynamics directly from video-derived supervisory signals.<n>We address this by shifting the predictive bottleneck from 2D image space to a lifted' 3D Volumetric Latent Space.<n>Unlike state-of-the-art hybrid models that rely on explicit classical simulators for training and/or inference, our architecture tracks material states implicitly within high-supervised V-JEPA features.
  arXiv  Detail & Related papers  (2026-06-24T22:06:35Z)
• PhyCo: Learning Controllable Physical Priors for Generative Motion [55.59209981836171]
  We present PhyCo, a framework that introduces continuous, interpretable, and physically grounded control into video generation.<n>Our approach integrates three key components: (i) a large-scale dataset of over 100K photorealistic simulation videos where friction, restitution, deformation, and force are systematically varied across diverse scenarios; (ii) physics-supervised fine-tuning of a pretrained diffusion model using a ControlNet conditioned on pixel-aligned physical property maps; and (iii) VLM-guided reward optimization, where a fine-tuned vision-language model evaluates generated videos with targeted physics queries and provides differentiable feedback.
  arXiv  Detail & Related papers  (2026-04-30T17:53:03Z)
• Interpreting Physics in Video World Models [28.27323417999283]
  A long-standing question in physical reasoning is whether video-based models need to rely on factorized representations of physical variables.<n>We present the first interpretability study to examine physical representations inside large-scale video encoders.<n>We find that modern video models do not use factorized representations of physical variables like a classical physics engine.
  arXiv  Detail & Related papers  (2026-02-04T15:19:19Z)
• Physics Steering: Causal Control of Cross-Domain Concepts in a Physics Foundation Model [13.900251746035012]
  We investigate the internal representations of a large physics-focused foundation model.<n>By injecting concept directions back into the model during inference, we can steer its predictions.<n>Our findings open new avenues for understanding and controlling scientific foundation models.
  arXiv  Detail & Related papers  (2025-11-25T19:40:22Z)
• LikePhys: Evaluating Intuitive Physics Understanding in Video Diffusion Models via Likelihood Preference [57.086932851733145]
  We introduce LikePhys, a training-free method that evaluates intuitive physics in video diffusion models.<n>We benchmark intuitive physics understanding in current video diffusion models.<n> Empirical results show that, despite current models struggling with complex and chaotic dynamics, there is a clear trend of improvement in physics understanding as model capacity and inference settings scale.
  arXiv  Detail & Related papers  (2025-10-13T15:19:07Z)
• PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation [53.06495362038348]
  Existing generation models excel at producing photo-realistic videos from text or images, but often lack physical plausibility and 3D controllability.<n>We introduce PhysCtrl, a novel framework for physics-grounded image-to-video generation with physical parameters and force control.<n> Experiments show that PhysCtrl generates realistic, physics-grounded motion trajectories which, when used to drive image-to-video models, yield high-fidelity, controllable videos.
  arXiv  Detail & Related papers  (2025-09-24T17:58:04Z)
• Latent Intuitive Physics: Learning to Transfer Hidden Physics from A 3D Video [58.043569985784806]
  We introduce latent intuitive physics, a transfer learning framework for physics simulation. It can infer hidden properties of fluids from a single 3D video and simulate the observed fluid in novel scenes. We validate our model in three ways: (i) novel scene simulation with the learned visual-world physics, (ii) future prediction of the observed fluid dynamics, and (iii) supervised particle simulation.
  arXiv  Detail & Related papers  (2024-06-18T16:37:44Z)
• Physics-Guided Human Motion Capture with Pose Probability Modeling [35.159506668475565]
  Existing solutions always adopt kinematic results as reference motions, and the physics is treated as a post-processing module. We employ physics as denoising guidance in the reverse diffusion process to reconstruct human motion from a modeled pose probability distribution. With several iterations, the physics-based tracking and kinematic denoising promote each other to generate a physically plausible human motion.
  arXiv  Detail & Related papers  (2023-08-19T05:28:03Z)
• Trajectory Optimization for Physics-Based Reconstruction of 3d Human Pose from Monocular Video [31.96672354594643]
  We focus on the task of estimating a physically plausible articulated human motion from monocular video. Existing approaches that do not consider physics often produce temporally inconsistent output with motion artifacts. We show that our approach achieves competitive results with respect to existing physics-based methods on the Human3.6M benchmark.
  arXiv  Detail & Related papers  (2022-05-24T18:02:49Z)
• Dynamic Visual Reasoning by Learning Differentiable Physics Models from Video and Language [92.7638697243969]
  We propose a unified framework that can jointly learn visual concepts and infer physics models of objects from videos and language. This is achieved by seamlessly integrating three components: a visual perception module, a concept learner, and a differentiable physics engine.
  arXiv  Detail & Related papers  (2021-10-28T17:59:13Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.