Recent advances in Artificial Intelligence have been driven by large-scale statistical modelsoperating over implicit representational spaces, with Large Language Models (LLMs) as the mostprominent example. While such systems demonstrate impressive capabilities across a range oftasks, their reliance on implicit representations and probabilistic continuation limits transparency,auditability, and control in safety-critical or regulated domains.This paper introduces Representation Models (RMs), a new class of AI systems in whichexplicit internal representations constitute the primary computational substrate. In an RM,reasoning, learning, and decision-making proceed through the construction, transformation, andstabilization of representations under declared constraints, rather than through modality-specificprediction alone. RMs separate validated knowledge from exploratory computation, enforcestructural invariants, and organize computation across hierarchical representational orders.We provide a precise architectural definition of RMs, clarify the distinction between knowledge,learning, reasoning, and evolution, and position RMs relative to language models, world models,and symbolic systems. This document serves as the foundational public introduction of theRM paradigm, focusing on architectural principles rather than performance benchmarks, andestablishing the conceptual framework required for subsequent instantiations and validationphases.
Simone Mazzoni (Sun,) studied this question.