The ability to understand and reason about cause and effect -- encompassing interventions, counterfactuals, and underlying mechanisms -- is a cornerstone of robust artificial intelligence. While deep learning excels at pattern recognition, it fundamentally lacks a model of causality, making systems brittle under distribution shifts and unable to answer ``what-if'' questions. This paper introduces the Hierarchical Causal Primitive Dynamic Composition Network (HCP-DCNet), a unified framework that bridges continuous physical dynamics with discrete symbolic causal inference. Departing from monolithic representations, HCP-DCNet decomposes causal scenes into reusable, typed causal primitives organized into four abstraction layers: physical, functional, event, and rule. A dual-channel routing network dynamically composes these primitives into task-specific, fully differentiable Causal Execution Graphs (CEGs). Crucially, the system employs a causal-intervention-driven meta-evolution strategy, enabling autonomous self-improvement through a constrained Markov decision process. We establish rigorous theoretical guarantees, including type-safe composition, routing convergence, and universal approximation of causal dynamics. Extensive experiments across simulated physical and social environments demonstrate that HCP-DCNet significantly outperforms state-of-the-art baselines in causal discovery, counterfactual reasoning, and compositional generalization. This work provides a principled, scalable, and interpretable architecture for building AI systems with human-like causal abstraction and continual self-refinement capabilities.
Figures & tables
Paradigm / Method
Causal Semantics
Compositionality
Neural-Symbolic
Self-Improvement
Physical Dynamics
Hierarchical Abstraction
Neural Causal Models (e.g., CausalVAE)
Explicit (Discrete)
Low
No
No
Limited
No
Modular Networks / GNNs
Implicit / Pre-defined
Medium
Partial
No
Sometimes
No
Differentiable Logic (e.g., LTN)
Explicit (Symbolic)
High
Yes
No
No
No
Meta-Learning (MAML, etc.)
No
Task-level
No
Yes (Task Adaptation)
Sometimes
No
World Models (e.g., Dreamer)
No
Low
No
No
Learned (Black-box)
No
Physics-Informed Neural Networks
No
Low
No
No
Explicit (PDE-based)
No
Table I : Positioning of HCP-DCNet against related paradigms.
Figure 1: System architecture of HCP-DCNet. Raw multimodal observations are processed by the Perception Engine into structured causal queries. The Primitive Library (Sec. III) supplies a hierarchical set of reusable causal primitives. The Dual-Channel Router (Sec. IV) dynamically selects and composes these primitives into a Causal Execution Graph (CEG) specification by fusing symbolic constraints from the Knowledge Graph with subsymbolic attentional patterns. This specification is executed by the differentiable Hybrid CEG Engine (Sec. V) to generate causal graphs and counterfactual predictions. The Meta-Evolution Controller (Sec. VI) closes the loop by analyzing performance from the Performance Log , proposing improvements to the primitive library and router, and updating the knowledge graph—enabling continual self-optimization. Arrows indicate the primary flow of data and control.
Method
CausalWorld (SHD ↓ )
SI-Blocks
C-H
Push
Stack
Chain
SHD ↓
CF-Acc ↑
CF-Acc ↑
Consist. ↑
NOTEARS [ 11 ]
1.2
3.8
6.1
4.5
0.62
0.58
0.71
CausalVAE [ 62 ]
0.9
2.5
4.8
3.2
0.68
0.65
0.78
DreamerV2+GNN [ 63 ]
0.8
2.1
3.9
2.8
0.71
0.63
0.75
CITRIS [ 65 ]
0.7
1.8
3.5
2.5
0.74
0.69
0.82
LTN [ 68 ]
1.1
3.1
5.5
1.9
0.79
0.72
0.88
Table II : Main results on CausalWorld (Push, Stack, Chain) and SI-Blocks tasks. Best results in bold , second best underlined . SHD ↓ is lower better, CF-Acc ↑ is higher better.
Figure 2: Experimental results comparing HCP-DCNet with state-of-the-art baselines across four key metrics. (a) Causal Graph Accuracy measured by Structural Hamming Distance (SHD, lower is better). HCP-DCNet achieves significantly lower SHD, indicating more accurate causal structure discovery. (b) Counterfactual Prediction Accuracy (CF-Acc, higher is better). HCP-DCNet shows superior performance in counterfactual reasoning tasks. (c) Compositional Generalization Score (CG-Score, higher is better) on novel combinations of causal primitives. HCP-DCNet demonstrates strong out-of-distribution generalization. (d) Inference Time in milliseconds (lower is better). While HCP-DCNet has higher computational cost due to its hierarchical routing, it remains practical for real-time applications. Error bars represent standard deviation over 5 random seeds. Ablation studies (HCP w/o Meta) highlight the importance of meta-evolution for optimal performance.
Causal discovery through experimentation and intervention is fundamental to robust problem solving. It requires not just updating beliefs within a fixed framework but revising the hypothesis space itself, a capacity current AI agents lack when evidence demands representations they have not previously constructed. We extend the blicket detector paradigm from developmental science to test this capacity in AI agents equipped with architectural scaffolding that targets hypothesis-space restructuring. Our compositional architecture has two discrete components: context graphs, which structure exploration as typed state machines, and dynamic behaviors, which monitor for evidence that the current hypothesis space is inadequate and expand it at runtime. Across 1,085 experimental trials, these components make orthogonal contributions: context graphs drive reasoning quality within the post-switch hypothesis space, accounting for 94% of the accuracy gain, while dynamic behaviors drive reasoning eligibility by detecting regime changes and preventing premature commitment to outdated hypotheses.
John Alderete, Sebastian Benthal, Connie Xu +1
Amigo AI · Simon Fraser University · International Computer Science Institute
We introduce CausaLab, a scalable environment for evaluating interactive causal discovery by LLM agents. Unlike prior evaluations, CausaLab evaluates both whether an agent can solve a problem using causal evidence and whether its answer is grounded in a faithful recovered causal mechanism. Each episode places an agent in a synthetic laboratory: it receives prior measurement records, intervenes on a manipulator crystal, and predicts the resonance frequency of a held-out reactor crystal governed by the same mechanism. The hidden data-generating process is a randomly sampled structural causal model (SCM), so success requires recovering both a causal graph and structural equations rather than recalling prior knowledge. Experiments show a persistent gap between prediction and mechanism recovery: in the purely observational 6-node setting, GPT-5.2-high reaches 92% task accuracy but only 0.471 all-edge F1. Mixed observation-intervention strategies improve structural fidelity, while pure intervention remains difficult even for strong agents. We identify premature stopping as a major weakness and show that consistency verification mitigates it. CausaLab therefore separates predictive success from causal understanding and exposes current LLM agents' limits as experimental causal reasoners.
Junlin Yang, Dylan Zhang, Xiangchen Song +7
Tsinghua University · University of Illinois Urbana-Champaign · Carnegie Mellon University +2
Efficiently navigating complex environments requires agents to internalize the underlying logic of their world, yet standard world modelling methods often struggle with sample inefficiency, lack of transparency, and poor scalability. We propose a framework for constructing symbolic causal world models entirely online by integrating continuous model learning and repair into the agent's decision loop, by leveraging the power of Meta-Interpretive Learning and predicate invention to find semantically meaningful and reusable abstractions, allowing an agent to construct a hierarchy of disentangled, high-quality concepts from its observations. We demonstrate that our lifted inference approach scales to domains with complex relational dynamics, where propositional methods suffer from combinatorial explosion, while achieving sample-efficiency orders of magnitude higher than the established PPO neural-network-based baseline.
Enrique Crespo-Fernandez, Oliver Ray, Telmo de Menezes e Silva Filho +1