Functional Model of Intelligence (FMI)

• FMI is a formal framework that defines intelligence as the observable capacity to construct, adapt, and apply internal models and reasoning methods in unknown settings.
• It quantifies intelligence via a scalar score based on knowledge, planning quality, learning speed, and reaction time, ensuring rigorous, substrate-neutral evaluation.
• FMI guides AGI development by enforcing recursive semantic coherence and multi-level alignment, which are vital for building robust and interpretable intelligent systems.

The Functional Model of Intelligence (FMI) defines intelligence strictly through the externally observable ability of a system to construct, adapt, and apply models, skills, and reasoning methods for achieving goals in variable and previously unknown environments. FMI is mathematically and architecturally formalized in recent literature spanning several diverse but convergent theoretical traditions, all emphasizing a substrate-neutral, continuously graded, and black-box-testable notion that isolates “what” intelligence does from “how” it is implemented. This paradigm enables principled comparison and engineering of both biological and artificial systems, and is positioned as foundational for AGI and robust alignment.

1. Mathematical Definition and Core Formalism

The FMI is defined by a scalar intelligence score that functionally evaluates a system based on its knowledge, planning competence, learning speed, and reaction time. In its canonical form (Sritriratanarak et al., 2023):

$$I_i = \alpha \left\lVert W^i_S - W^*_S \right\rVert_w + \frac{ \beta \left\lVert W^i_{\mathrm{goal}} - (A'(W^i_S)\Rightarrow W^i_{S'}) \right\rVert_w }{ \gamma t_{A'} + \delta \frac{\partial W^i_S}{\partial D} + \varepsilon \frac{\partial f(W^i_S, W^i_{\mathrm{goal}})}{\partial D} - \zeta \frac{\partial t_{A'}}{\partial D} }$$

where:

• $W^i_S$ is the system's internal world-model, decomposed into abstract models ($M^i$), actions ($A^i$), objects ($O^i$), and relations ($R^i$).
• $W^*_{S}$ is the observer-accessible “true” world state.
• $W^i_{\mathrm{goal}}$ encodes the internal goal representation.
• $A'$ is the plan produced by the system, with $A'(W^i_S)\Rightarrow W^i_{S'}$ as the predicted post-plan world state.
• $f(W^i_S, W^i_{\mathrm{goal}})$ is the planning function (policy generator).
• $t_{A'}$ is the real time to compute $A'$; $D$ is new data or experience.
• $\left\lVert \cdot \right\rVert_w$ is a world-norm, weighted over the errors in models, actions, objects, and relations.
• Parameters $\alpha, \beta, \gamma, \delta, \varepsilon, \zeta$ set trade-offs between knowledge, planning quality, speed, and learning rates.

This formalism admits continuous scoring, is grounded in experimenter-observable variables, and is agnostic to internal mechanisms (Sritriratanarak et al., 2023).

2. Distinction from Related Concepts and Theoretical Motivation

FMI is explicitly differentiated from several commonly conflated or anthropomorphic notions (Sritriratanarak et al., 2023, Pfister, 10 Mar 2025):

• Sensations: Functional triggers (e.g., pain signals) do not require reasoning, and are not intelligence.
• Autonomy: Acting without oversight is orthogonal; intelligence is not agency per se.
• Skill: Narrow proficiency can be the output of brute force or lookup and is not intelligence unless the skill can be constructed for new settings.
• Sentience: Self-modeling or consciousness is neither necessary nor sufficient for intelligence within FMI.
• Intentionality: FMI does not presuppose or require any “aboutness” or qualia—representations are physical states serving only functional roles (Pfister, 10 Mar 2025).

Intelligence, in FMI, is the capacity to construct novel skills (functions from representations to actions) for previously unknown contexts and under indirect or incomplete perception, evaluated by their effectiveness in achieving specified goals. This positioning aligns with functionalist and naturalistic interpretations (Pfister, 10 Mar 2025).

3. Functional Decomposition and Recursive Coherence

Within FMI, intelligence is decomposed into three core capabilities (Sritriratanarak et al., 2023):

• Knowledge: Construction of internal models accurately reflecting relevant objects and relations.
• Reasoning: Both inference (logical deduction, induction, abduction) and planning (action sequence search).
• Learning: Adaptive improvement of knowledge and reasoning speed/quality with respect to new experience $D$.

A necessity for large-scale and multidomain intelligence is recursive semantic coherence across stacking reasoning levels. The Recursive Coherence Principle (RCP) states that only architectures equipped with a minimal set of internal functions—evaluation, modeling, stability, adaptation, decomposition, and bridging—preserve semantic alignment under recursion (Williams, 18 Jul 2025). The FMI is formalized as the unique operator algebra capable of satisfying the RCP at all reasoning orders. This guarantees:

• Coherence-checking: Every composite transformation on a unified conceptual space $C_N$ is audited for semantic preservation.
• Repair: Incoherence (misalignment, instability, hallucination) is detected and remediated via compositional correction routines.
• Bridging and Decomposition: Enables cross-domain reasoning and modular repair (Williams, 18 Jul 2025).

The absence of any FMI primitive provably leads to misalignment, semantic drift, and failure of scalable inference and coordination.

4. Representationalist and Constructivist Foundations

FMI adopts a representationalist, constructivist approach, whereby (Pfister, 10 Mar 2025):

• World models $\mathcal M$ are repositories of representations ($r_i$), skills ($s_j$), and inferences, all derived from indirect, noisy, and ambiguous percepts.
• Inference methods—deduction (certain but non-creative), induction (generalization), and abduction (hypothesis generation and vocabulary extension)—expand and refine $\mathcal M$.
• Abstraction and classification regulate complexity by grouping and reducing redundancy in internal representations.
• Viability is defined as the goal-achievement probability or expected utility of a representation; only representations supporting actionable policies ($\pi_r$) above a utility threshold are meaningful.

Meaning is thus functionally ascribed—“$r$ means ‘food’” only if the agent’s probability of goal $G$ rises when $r$ is interpreted as food. This paradigm avoids grounding meaning or value in consciousness or intentional experience (Pfister, 10 Mar 2025).

5. Architectures, Layered Representations, and Function Alignment

FMI architectural models frequently instantiate multi-level representational hierarchies (e.g., sensorimotor, symbolic) coupled via bidirectional prediction and encoding/decoding mechanisms (Xia, 27 Mar 2025). “Function alignment” is established when distinct representations at different abstraction layers (e.g., subsymbolic $\boldsymbol x$ and symbolic $\boldsymbol z$) are aligned temporally and referentially to the same ground-truth sequence ($\boldsymbol y$):

• Bidirectional auto-regressive coupling links transitions in each layer to the state of the other.
• Layered update equations combine horizontal (self-transition), vertical (encoding/decoding), and diagonal (cross-layer, cross-time) information flow.
• Bounded interpretability: Any transformation between layers incurs irreducible approximation error $\epsilon$, unifying the phenomena of bounded rationality, symbol grounding, and analogy into a single representational approximation constraint (Xia, 27 Mar 2025).

The “Isomorphic Alignment Theorem” shows that such functionally aligned multi-layer systems are mathematically isomorphic to higher-dimensional unified agents. For design, FMI prescribes enforcing alignment by integrating bidirectional prediction heads and minimizing a joint loss spanning all coupling terms.

6. Measurement, Practical Challenges, and Theoretical Limits

Quantifying FMI and applying it in real settings faces critical practical obstacles (Sritriratanarak et al., 2023, Pfister, 10 Mar 2025):

Challenge	Source Reference	Significance
Observer’s incomplete $W^*$	(Sritriratanarak et al., 2023)	External measurement error limits evaluation of knowledge accuracy
Outcome evaluation complexity	(Sritriratanarak et al., 2023)	Requires objective access to system predictions and actual results
Ontological boundary fuzziness	(Sritriratanarak et al., 2023)	Novel abstractions confound external assessment of internal model
Derivative/learning rate estimation	(Sritriratanarak et al., 2023)	Controlled experiments needed to isolate adaptation speed
Parameter calibration	(Sritriratanarak et al., 2023)	Weighting importance of cognitive facets is application-dependent
No Free Lunch theorems	(Pfister, 10 Mar 2025)	Assumptions about world regularities are necessary for competence

These difficulties underscore FMI's epistemological and operational demands: the need for rigorous experimental control, robust norm definitions, and conscious tradeoffs in design parameters.

A crucial theoretical constraint—arising from the No Free Lunch theorems—is that any FMI-based system must embed biases matched to the regularities of its environment to surpass random search. Generality is thus inherently limited: greater performance in structured domains demands more specific priors, and vice versa (Pfister, 10 Mar 2025).

7. Applications, Alignment, and Systemic Impact

FMI’s implications are broad:

• AGI design: It serves as a mathematically grounded and naturalistic foundation for constructing substrate-neutral, continuously scalable intelligent systems (Pfister, 10 Mar 2025, Williams, 18 Jul 2025).
• Alignment: Ensures structural, not just behavioral, alignment via recursive coherence. Omitting any FMI primitive (evaluation, modeling, stability, adaptation, decomposition, or bridging) provably results in misalignment, hallucination, or collapse at scale (Williams, 18 Jul 2025).
• Interpretability and analogy: Bounded mappings between layers explain and constrain rationality, symbol grounding, and transfer, forming the basis for interpretable multi-level models (Xia, 27 Mar 2025).
• Practical blueprint: For system designers, FMI prescribes a joint architecture comprising distinct representation modules, enforceable alignment, and a compositional, recursively monitored reasoning engine. This scaffolds both rapid, heuristic (System 1) and deliberative, compositional (System 2) cognition.

This summary draws on (Sritriratanarak et al., 2023, Xia, 27 Mar 2025, Pfister, 10 Mar 2025), and (Williams, 18 Jul 2025) for all definitions, formalizations, and architectural prescriptions of the Functional Model of Intelligence.