X-FOREST Nets: A Hybrid State-Space Liquid Neural Engine with Structural Causal Memory for AGI

Author: Sudhan Regmi
Affiliation: Infi Web Tech Pvt. Ltd.

---

Abstract

Current autoregressive architectures rely on discrete token matching ($P(Y \mid X)$), incurring quadratic compute complexity $O(N^2)$, physical trajectory hallucinations, and catastrophic forgetting. This paper presents X-FOREST Nets, a causal, continuous dynamical world engine designed for Artificial General Intelligence (AGI).

X-FOREST decouples representation into a hybrid latent loop that bidirectionally couples Selective State Space Networks (SSN) with continuous-time Liquid Neural Networks (LNN). This engine ($\text{SSN} \rightleftharpoons \text{LNN}$) achieves linear time complexity $O(N)$ for long sequence histories while preserving fluid, continuous-time physics governed by Lagrangian mechanics ($\frac{dh}{dt}$). To eliminate factual and causal hallucinations, the latent state actively projects execution traces into a Differentiable Structural Causal Model (SCM) governed by NOTEARS continuous optimization and a real-time Knowledge Graph synchronization pipeline (Neo4j + GraphRAG).

Key Architectural Pillars

• Linear Context Scaling: Replaces quadratic attention with Selective State Space Networks ($O(N)$).
• Physical Grounding: Integrates Liquid Neural Networks constrained by Euler-Lagrange equations ($\mathcal{L}_{\text{Euler-Lagrange}} = 0$) to eliminate motion hallucinations.
• Causal Reasoning: Employs continuous NOTEARS optimization ($\text{Tr}(e^{W \circ W}) - d = 0$) to enable true interventional $P(Y \mid \text{do}(X))$ reasoning.
• Factuality Preservation: Synchronizes generated concept triplets into a relational Knowledge Graph pipeline.

Citation

If you reference or build upon this work, please cite:

@article{regmi2026xforest,
  title={X-FOREST Nets: A Hybrid State-Space Liquid Neural Engine with Structural Causal Memory for Artificial General Intelligence},
  author={Regmi, Sudhan},
  journal={Infi Web Tech Research},
  year={2026}
}