🎐 @saturdayrobotic Robotics & World Models Reading Club #13#, @huang_biwei (Assistant Professor at @UCSanDiego & Founder of @AetherLab_AI) delivered a deep technical vision for what a true Causal World Model (CWM) should be—and why today's frontier AI systems still fall short. Current paradigms: 🔹 LLMs (GPT, Claude, Gemini, π): language correlation models with weak grounding in physical interventions. 🔹 Video generators (Sora, Wan, Veo 3): high-fidelity pixel prediction but limited mechanism modeling, counterfactual reasoning, or causal guarantees. 🔹 JEPA-style models (V-JEPA, AMI): promising representation learning, but not yet fully causal world models. Key diagnosis: ❌ Modeling correlations ✅ Modeling how latent causes generate observations and how the world evolves under actions Causal World Model learns: 📥 Inputs: • Observations • Interventions / Actions • Temporal Information • Context / Environment • Other Agents / Entities ⚙️ Core Components: 1️⃣ Causal Feature Representation Learning Encoder → latent causal variables Z, recovering disentangled causal factors rather than merely compressing observations. 2️⃣ Causal Structure Learning Learn causal graph G: • Directed arrows → causal effects • Dashed links → latent confounders • Explicit latent causal variables and hidden drivers 3️⃣ Causal Dynamic Learning Temporal causal model over t → t+1 → t+2 with intra-time and inter-time causal dependencies. Benefits: ✅ Generalization under distribution shift ✅ Precise forecasting & action control ✅ Data-efficient and model-size-efficient learning ✅ Intervention & counterfactual reasoning ✅ Discovery of hidden mechanisms and new knowledge A recurring principle across tabular data, questionnaires, images, videos, time series, and genomics: 👉 Representation learning should recover latent factors and causal relations among them—not just compress observations. Examples: 📊 Tabular/questionnaire data → latent DAGs with causal variables and latent factors U1–U5 (NeurIPS 2022/2023, ICLR 2024, ICML 2024). 🎥 Images/videos → compositional concept graphs ("chicken eating rice" + "peacock" → "peacock eating rice"), hierarchical concepts, temporal pose graphs z₁,₁…z₃,₆ (CVPR 2026, ICLR 2025/2022, NeurIPS 2024/2022). 📈 Time series → trend, seasonal, cyclical, noise decomposition. 🧬 Genomics → motifs, binding sites, regulatory features. Strong critique of current video world models: ❌ Precise Action Control (WorldGym) ❌ Object Consistency (UniSim) ❌ Physical Consistency (CogVideoX) Beyond scaling, world models need explicit causal structure. Mathematically, the framework jointly models: • Observation function oₜ • Reward function rₜ • Hierarchical latent states sᵢ,ₜ • State dynamics • Action variables aₜ • Structural relation masks Dₛ(·,i) Empirical results: 🏃 Walker2d-v2 & Humanoid-v2: Causal Dynamic Learning (CDL) consistently outperforms Curiosity, CID, and ASR in: • Lower next-state prediction error • Higher policy returns For distribution shifts, the framework introduces: 🔹 Domain factors c 🔹 Structural adaptation parameters θₖ Key insights: • Graphs compactly encode what changed and where • Not all state dimensions matter for action prediction • Not all θ need adaptation • Factors can be adapted independently 🎮 CoinRun experiments further show that new latent variables and structural relations (adding new state dimensions such as s₄) can be introduced while preserving causal factorization, enabling efficient adaptation to new environments, mechanics, and visual styles. World Models = Causal Feature Learning + Causal Graph Discovery + Causal Dynamics + Factorized Distribution Shift Modeling Only then can AI achieve: 🔮 Counterfactual simulation 🎯 Intervention reasoning ⚡ Data-efficient adaptation 🔬 Discovery of previously unknown mechanisms This is not simply scaling larger models—it is a proposal to move from predicting what happens next toward understanding why the world changes.