Causal Driven World Model
Observe. Intervene. Learn.
A next-generation foundation model that does more than passively perceive data — it builds causal insights, makes informed decisions, and continually refines its actions through intervention, feedback, and real-world impact.
Vision
Beyond next-token prediction
Toward structural consequence computation
The current AI paradigm is data-hungry and correlation-bound. Most public corpora are exhausted, and bigger models memorize associations without uncovering the mechanisms that produce them.
By acting to learn — designing interventions, gathering feedback, and updating a structured world model — we break that ceiling.
Architecture
The observe–intervene–learn loop
Three modules that close the cycle from passive perception to active causal learning.
Observe
Perception modules learn a structured world model — a latent causal graph over entities, states, and actions.
Intervene
A causal planner selects low-risk, high-information interventions to maximize learning per interaction.
Learn
Updates beliefs from interventional feedback, closing the observe–act–learn cycle.
Observe
Perception modules learn a structured world model — a latent causal graph over entities, states, and actions.
Intervene
A causal planner selects low-risk, high-information interventions to maximize learning per interaction.
Learn
Updates beliefs from interventional feedback, closing the observe–act–learn cycle.
Impact
What a causal world model unlocks
Data Efficiency
Learning from interventions rather than brute-force memorization, breaking the "more data = better model" ceiling.
OOD Generalization
Structural understanding transfers across distributional shifts where pattern-matching fails.
Intervention-ready Planning
The model reasons about "what if I do X?" rather than only "what have I seen before?"
Explainable Decisions
Causal structure provides transparent, auditable decision pathways grounded in mechanism.
Publications
Related research
Learning World Models with Identifiable Factorization
Yu-ren Liu, Biwei Huang, Zhengmao Zhu, Honglong Tian, Mingming Gong, Yang Yu, Kun Zhang
NeurIPS 2023
Towards Generalizable Reinforcement Learning via Causality-guided Self-adaptive Representations
Yupei Yang, Biwei Huang, Fan Feng, Xinyue Wang, Shikui Tu, Lei Xu
arXiv preprint arXiv:2407.20651 (2024)
Modeling Unseen Environments with Language-guided Composable Causal Components in Reinforcement Learning
Xinyue Wang, Biwei Huang
arXiv preprint arXiv:2505.08361 (2025)