Skip to main content
Explicit Representation, Controlled Reasoning, and the Limits of Language‑Centric AI
Simone Mazzoni — RCUBEAI Research Lab (Paris, France) — December 2025

​
Abstract

Large Language Models (LLMs) are powerful but remain language‑centric and rely on implicit representations learned from data. This creates well-known limitations in regulated or safety‑critical contexts: opacity, limited auditability, and difficulty enforcing invariants. This paper introduces Large Representation Models (LRMs): AI systems whose primary object of computation is explicit internal representations. LRMs:
  • decouple language from reasoning,
  • separate validated knowledge from exploratory computation,
  • enforce structural invariants and closure constraints,
  • organize cognition across hierarchical representational Orders,
  • and support controlled representational growth (promotion and scaling).

​
Executive summary

Core shift

Intelligence is treated as representation construction + validation, not token continuation.

Governance

Explicit constraints, thresholds, and invariants enable auditability and controlled reasoning.

Scaling path

Complexity is absorbed via vertical structure (Orders), not only parameter count.

​
1. Why LRMs

LLMs treat language as both:
  1. the primary input modality, and
  2. the internal substrate of reasoning.
This architectural flattening makes it difficult to:
  • preserve hard constraints,
  • expose validity thresholds,
  • prevent competence collapse (hallucination as structural failure),
  • and certify behavior under domain rules.
LRMs propose a structural redesign: explicit, hierarchical representational computation.

​
2. Representation‑centric intelligence

​
2.1 Language‑centric vs representation‑centric

An LRM does not ask:
“What is the most probable continuation of this text?”
It asks:
“What representation of the domain is consistent, stable, and valid under applicable constraints?”

​
2.2 Orders and hierarchical decomposition

Problems are decomposed across Orders of representation:
  • lower Orders: concrete facts, syntax, local relations,
  • higher Orders: abstractions, meta‑constraints, cross‑domain coherence.
This reduces category errors (e.g., resolving normative conflicts with statistical heuristics).

​
3. Four cognitive modes

Validated, stationary representations available at a moment in time.
In LRMs, knowledge is explicitly represented and auditable.

​
4. Implicit vs explicit representation

​
LLMs (implicit)

  • reasoning occurs in a learned embedding geometry,
  • constraints are distributed across parameters,
  • thresholds and invariants are not isolated or controllable.

​
LRMs (explicit)

  • constraints and thresholds are represented directly,
  • knowledge is separated from exploration,
  • propositions are evaluated against declared invariants,
  • outputs are “committed representational objects” with traceable validity conditions.

​
5. Competence stabilization and hallucination prevention

LLM failure modes can be explained by flat architecture:
  • higher-level inference can override lower-level constraints silently.
LRMs enforce competence stabilization:
  • knowledge validated at an Order becomes stationary at that Order,
  • for higher Orders, that stabilized structure becomes competence (non‑negotiable constraints),
  • contradictions become explicit non‑closure instead of silent violations.

​
6. Orders as computational spaces (intrinsic multi‑agenticity)

In LRMs, an Order is a coherent computational space characterized by:
  • admissible representational objects,
  • admissible operations and transformations,
  • validity constraints and closure conditions,
  • resource bounds.
Because each Order has its own constraints and dynamics, the architecture is intrinsically multi‑agentic (without heuristic tool-calling): each Order acts as a specialized reasoning unit with explicit interfaces to adjacent Orders.

​
7. Vertical dynamics: promotion vs scaling

​
7.1 Two causes of non‑closure

  • Abstraction deficit: representation is not expressive enough.
  • Precision deficit: representation is expressive enough but under‑resolved (wide intervals, missing fine constraints, boundary conditions).

​
7.2 Two vertical operators

  • Promotion (Vertical Ascension) resolves abstraction deficit by introducing a higher Order.
  • Scaling (Vertical Descension) resolves precision deficit by decomposing into sub‑orders (refinement spaces).
Scaling creates a resolution tree: closure may require descending to refine local constraints, then reintegrating upward as stabilized competence.

​
8. Knowledge vs competence

Knowledge: what is validated and operational within an Order.
Competence: what is inherited by an Order from lower Orders, functioning as fixed constraints. The same structure can be:
  • knowledge at the Order where it is validated,
  • competence for higher Orders.

​
9. Canonical definition (formal)

An LRM is an artificial cognitive system whose primary computation operates over explicit, structured, hierarchical representational spaces. Formal tuple: 
L = (O, {R_N}, {C_N}, {I_N}, D)
Where:
  • O: partially ordered set of Orders (indexed by N)
  • R_N: representational space at Order N
  • C_N: constraints, invariants, admissible operations governing R_N
  • I_N: interfaces linking R_N to adjacent Orders and external modalities
  • D: dynamical operators (reasoning, learning, promotion, scaling)

​
10. Positioning LRMs in the AI landscape

​
11. Architectural instantiation (per Order)

Within each Order N, three components interact:
  • Substrate Φₙ: stabilized representations (knowledge + inherited competence)
  • Projection space Ψₙ: dynamic exploration and reasoning
  • Interface Wₙ: observation, closure evaluation, representational commit

​
Wₙ as observer and closing gate

Wₙ:
  • evaluates candidate structures from Ψₙ against Φₙ,
  • decides closure under constraints of the Order,
  • commits validated representational objects to Φₙ,
  • emits structured non‑closure when closure cannot be achieved.

​
12–13. Orchestration and multi‑clock computation

LRMs do not assume a single global execution timeline:
  • each Order has an Order‑relative “clock” (its own progression regime),
  • coordination is achieved via scoped synchronization windows and consistency conditions,
  • closure decisions remain local to Wₙ (not a central scheduler).
In mature LRMs, orchestration is presented as an emergent property of representational laws and hierarchical constraints, though early implementations may use explicit orchestration layers pragmatically.

​
14. LRMs and safe AGI (glass‑box requirement)

LRMs are positioned as a candidate foundation for safe AGI because they:
  • make reasoning and representational change traceable,
  • enforce constraints hierarchically (constraints cannot be bypassed by higher‑level reasoning),
  • treat non‑closure as a signal for governed evolution rather than forced outputs.

​
15–16. Scope, grounding, and roadmap

This is an architectural definition paper:
  • it avoids internal algorithmic implementations and engineering details,
  • it states theoretical grounding in the broader R3 program, while remaining evaluable independently at the architecture level,
  • it reports a completed foundational validation phase and outlines a roadmap toward mature LRM properties (including autonomous promotion/scaling and intrinsic orchestration).

​
17. Conclusion

LRMs reorganize artificial cognition around:
  • explicit, hierarchical representation,
  • controlled reasoning with competence preservation,
  • governed evolution driven by non‑closure,
  • and glass‑box auditability.

​
Suggested citation

Mazzoni, S. (2025). Large Representation Models (LRM): Explicit Representation, Controlled Reasoning, and the Limits of Language-Centric AI. RCUBEAI Research Lab.