CausalPhysics: Unifying Semantic Reasoning, Physical Dynamics, and Counterfactual Simulation in World Models

Current world models fragment physical intelligence into separate pipelines. Vision language models (VLMs) excel at semantic tasks but struggle with causal physical reasoning: on our CAUSALPHYSICS-BENCH evaluation, GPT-4V answers only 21.9% of counterfactual physics queries correctly. Video generators produce realistic frames but understand little physics: Sora attains 24.1%, Runway Gen-3 23.2%, and VideoPoet 21.4% on Physics-IQ (Motamed et al., 2025). Model-based reinforcement learning (MBRL) systems operate in narrow domains and lack semantic grounding. We present CAUSALPHYSICS, a single architecture that bridges these gaps with three tightly coupled modules: (1) a Semantic-Physical Encoder (SPE) that fuses DINOv2 vision tokens with frozen LLaMA-2 language representations through cross-attention; (2) a Causal Graph Induction Module (CGIM) that discovers a differentiable structural causal model from video, supporting Pearl’s do-operator and counterfactual queries; (3) a Physics-Constrained Dynamics Network (PCDN) that propagates states through the learned causal graph while enforcing differentiable conservation-law constraints. On the official Physics-IQ v1.0 toolkit, CAUSALPHYSICS scores 46.8 ± 0.9—a 47% relative gain over V-JEPA 2 (31.8 ± 1.4) and roughly double Sora (24.1). Causal consistency reaches 71.3 ± 1.2% on CAUSALPHYSICS-BENCH versus 21.9 ± 0.8% for GPT-4V ($p<0.001$, paired t-test, 3 seeds). Out-of-distribution (OOD) generalization improves by 20.2 percentage points over the strongest baseline