This work introduces Unified Relational Intelligence (URI-v2), a novel artificial intelligence framework designed to move beyond statistical correlation toward causal unsupervised learning. Unlike conventional machine learning systems—particularly Large Language Models—which rely on pattern recognition within large datasets, URI proposes a fundamentally different paradigm in which causality is embedded directly into the mathematical structure of the learning architecture.The framework is grounded in Pattana-Relational Dynamics (PRD), where the 24 universal causal conditions (Paccayas) are mapped onto the SU(5) Lie algebra, forming a structured relational manifold that governs how an AI system learns and represents knowledge. Within this manifold, causal relations are encoded as algebraic operators rather than probabilistic correlations, allowing the system to construct world models through the discovery of relational dynamics.At the core of the architecture is the URI Relational Learning Engine (URI-LE), which replaces traditional attention mechanisms with a causally grounded relational interaction layer. Instead of learning statistical token dependencies, the system learns a set of relational operators that describe how states transform according to the algebraic structure of SU(5). These operators form the basis of a causal state representation defined in a relational Hilbert space, where system states evolve according to a relational Hamiltonian consistent with the underlying symmetry structure.Learning in URI-LE is driven by symmetry preservation and algebraic consistency, ensuring that all learned transformations respect the commutation relations of the SU(5) algebra. A dedicated Paccaya Filter evaluates candidate outputs to ensure causal validity, rejecting outputs that violate relational constraints. Long-range dependencies are modeled using an Upanissaya weighting mechanism, which captures decisive causal influences across temporal scales.The framework further introduces a Relational Attention mechanism, extending transformer architectures by embedding causal operators directly into attention computations. This allows the system to prioritize tokens or states based on genuine causal relationships rather than statistical co-occurrence. Multi-head relational attention enables the model to capture multiple categories of causal interaction simultaneously.URI-v2 provides a complete algorithmic pipeline for causal unsupervised learning, including relational encoding, operator discovery, symmetry-preserving training objectives, and Riemannian optimization on Lie groups. The framework supports applications ranging from causal reasoning in natural language processing to multi-modal world modeling.By integrating concepts from causal machine learning, geometric deep learning, and mathematical physics, URI-v2 represents a step toward artificial intelligence systems capable of discovering causal structure directly from raw data. This approach offers a potential pathway toward AI systems that learn through relational understanding rather than statistical approximation, addressing key limitations of current large-scale machine learning models.
Myomin Aung (Mon,) studied this question.