Interpretation as Linear Transformation: A Cognitive-Geometric Model of Concepts and Meaning

AI summary

Overview

• Research area: Artificial intelligence theory, formal epistemology, cognitive science, and conceptual-space modeling (arXiv:2512.09831v2 [cs.AI], sole author Chainarong Amornbunchornvej, National Electronics and Computer Technology Center, Thailand).
• Technical level: Intermediate. The paper is largely conceptual and philosophical, but it is built on linear algebra (vector spaces, linear maps, null spaces, norms, convex hulls) and includes formal theorems in appendices.
• Scope: The paper proposes a single geometric formalism in which every agent has a personalized vector space of values, concepts are vectors inside it, and communication is a linear transformation between agents' spaces.

What This Paper Is About

Most models of communication assume that concepts and beliefs are propositional items that can be transmitted intact, so misunderstanding is blamed on bad information, noise, bias, or faulty reasoning. This paper argues the real barrier is structural: agents differ in the dimensions through which they carve up and evaluate meaning, so a message is not copied but transformed into a new representational basis. The goal is to give a precise algebraic account of when a concept survives that transformation, when it is distorted, and when it is annihilated (concept "death").

Key Contributions

1. A geometric ontology of evaluative concepts. Each agent is modeled as a finite-dimensional value space whose basis vectors are the agent's fundamental evaluative dimensions. Concepts are structured vectors called abstract beings; a belief is simply an abstract being that an agent has endorsed and that is motivationally active. Meaning is shifted from propositional content to transformational structure.
2. A structural account of intelligibility and influence. Communication is a linear interpretation map T_{A→B}: 𝒱_A → 𝒱_B. Whether a concept survives transmission depends on whether it avoids the map's null space, which yields geometric criteria for intelligibility, distortion, miscommunication, and concept death (a being dies when it falls into the null space of every agent's interpretation map). The paper names structural reachability as a precondition for influence.
3. A unified model of motivational dynamics. Motivational gradients live in the same space as concepts, so adopting an abstract being changes an agent's goal vector and action tendencies. This supports formal accounts of goal drift, value convergence, and alignment with influential agents.
4. A model of innovation and value-space expansion. When an innovator occupies a vector outside the convex hull of a group's current valuations, they introduce genuinely new evaluative directions that the group could not generate endogenously — expanding the group's representational geometry rather than interpolating inside it.

A fifth framing element runs through the paper: the No-Null-Space Leadership Condition, described as a central result that characterizes leadership as a property of representational reachability rather than persuasion or authority.

Main Findings

• Null spaces as the mechanism of cognitive blindness: A concept that falls into the null space of T_{A→B} is invisible to the recipient. The paper's worked example uses x_A = [1.0, 0.5, 0.0] (creativity, social utility, financial reward) and a map that zeroes the creativity dimension, giving x_B = [0, 0.5, 0]; a purely creative act z_A = [1, 0, 0] maps to 0. This formalizes value misalignment as total perceptual omission rather than mere disagreement.
• Three consistency conditions define genuine influence: forward consistency (the transmitted being keeps its representational structure under T_{A→B}), backward consistency (B's representation can be approximately mapped back to A's, enabling mutual intelligibility), and valuation consistency (the concept retains or gains subjective value in the recipient's space, preventing motivational collapse).
• Local coherence (Theorem 1): If each of two agents can approximately reconstruct the other's representation, repeated communication preserves concept stability — a formal version of representational intersubjectivity across differing cognitive geometries.
• Understanding as angular proximity: In the mutual-understanding example, a goal vector x = [0.7, 0.8] (education, social impact) is projected by two different agents' maps into T_A(x) = [1.1, 0.48, 0] and T_B(x) = [1.08, 0.14, 0.15]; the cosine similarity between these images is approximately 0.96, which the paper reads as high epistemic compatibility.
• Valuation can diverge dramatically under a change of perspective: Applying T_{A→B} = diag(0.3, 0.5, 1.2) to x_A = [0.9, 0.6, 0.2] gives x_B = [0.27, 0.3, 0.24]. With a component-sum valuation function, Val_A = 1.7 while Val_B = 0.81 — the same goal is valued at less than half as much by the second agent.
• Motivation is a gradient in value space: M_i = g_i − x_i. For a current state x_i = [0.4, 0.3, 0.2] and goal g_i = [0.9, 0.6, 0.4], the gradient is [0.5, 0.3, 0.2], i.e., the strongest drive is toward the first dimension (knowledge), then reputation, then income.
• Concepts act as directional attractors: Alignment between a concept b and a motivational gradient is measured by a dot product. For b = [0.6, 0.8, 0.0] ("Service to others brings fulfillment") and M_i = [0.5, 0.5, 0.1], the alignment score is 0.6·0.5 + 0.8·0.5 + 0·0.1 = 0.7.
• Successful influence preserves rank-order structure and magnitude: For b = [1.0, 0.6] (freedom, collective benefit), T_A(b) = [1.3, 0.48, 0] and T_B(b) = [1.14, 0.16, 0.2]; Val_A ≈ 1.38 and Val_B ≈ 1.17, with cosine similarity near 1 — structurally different but motivationally resonant.
• Epistemic compatibility as overlapping, differently weighted subspaces: Interpreting "Freedom" c = [1.0, 0.5] through two different maps yields [1.0, 0.25, 0] and [1.0, 0, 0.25] — both agents find the concept valuable, but one dimension (achievement) is annihilated in one space while remaining small in the other, producing systematic misunderstanding.
• Trans-agent identity is graded and asymmetric: Two vectors count as the same abstract being to the extent that maps in both directions transport them into approximate images of one another, within a round-trip distortion ε. Identity is therefore governed by ε, need not be symmetric, and is grounded in the algebra of the maps rather than any external criterion of sameness.
• Influence in networks is constrained by composition, not just topology: In a multi-agent network, a being propagates if it survives the cognitive filters at each interpretive link along a connected path. This contrasts with probabilistic diffusion models (e.g., Kempe et al. 2003; Jackson and Yariv 2007) by adding representational constraints absent from those frameworks. The illustration uses b = [1.0, 0.7] in ℝ² ("Decentralized knowledge is more resilient") originating with agent A_1, but the details of the outcome are not reported in the available content.
• A shared-basis hypothesis explains when interpretation maps are well-behaved: Agents in a common occupational, disciplinary, or social practice acquire the same basis vectors through repeated joint activity. On that shared task subspace 𝒰 ⊆ 𝒱_A, the round-trip operator T_{B→A} ∘ T_{A→B} satisfies local coherence for small ε. Null spaces become consequential at the boundaries of shared bases — cross-occupational, cross-disciplinary, or cross-cultural exchange — so an engineer's reasoning about algorithmic complexity, a philosopher's about transcendental idealism, or a clinician's about autoimmune etiology will be largely annihilated for a receiver lacking the corresponding basis.
• Linearity is defended as a first-order approximation: Any smooth mapping between cognitive spaces can be locally approximated by its Jacobian, just as a Taylor expansion linearizes nonlinear functions, so the linear treatment captures projection, distortion, and annihilation while remaining analytically tractable.
• Philosophical connections are made algebraic: Dennett's intentional stance becomes the selection of an interpretation map that renders one agent's abstract beings coherent in another's value space; Clark's common ground becomes the shared task subspace on which T is near-invertible; Floridi's Level of Abstraction is paralleled by abstract beings defined relative to an agent's basis; and Longino's emphasis on perspectival divergence is recast as structural incompatibility of value spaces.
• The framework explicitly positions itself against prior work: Conceptual-space models (Gärdenfors 2004, 2014) typically presuppose shared or comparable dimensions; neuroscience representational-geometry work (Kriegeskorte and Kievit 2013; Greco et al. 2024; Wei and Woodford 2025) remains largely perceptual or task-specific; and logical/symbolic belief models (Levesque 1984) abstract away from representational transformation. The claimed gap is a structural criterion for concept survivability under heterogeneous cognitive architectures.

Methodology in Plain English

The paper is theoretical: it defines a mathematical vocabulary and then derives consequences from it, rather than running experiments on data.

1. Give each agent a private vector space. An agent's value space is a finite-dimensional vector space whose basis vectors stand for the evaluative dimensions that matter to that agent — moral priorities, pragmatic considerations, identity concerns, epistemic preferences, and so on. Its dimension count is whatever matches the agent's interpretive structure.
2. Represent concepts as vectors. Rather than treating a concept as a sentence-like proposition, it becomes a vector in that space. The vector's norm reflects how important the concept is to the agent; its direction reflects how the concept is internally organized.
3. Model communication as a linear map. When agent A expresses a concept to agent B, the concept is projected through T_{A→B}, which can rotate, rescale, or delete components. Whatever lies in the map's null space becomes unintelligible.
4. Add motivation as a second vector in the same space. An agent's current state and desired goal state are both points; their difference is a motivational gradient pointing in the direction the agent is inclined to act.
5. Define survival, coherence, and death algebraically. A concept survives if it avoids the relevant null spaces and retains comparable magnitude and directional structure. Two agents achieve local coherence if each can approximately reconstruct the other's representation. A being "dies" when no agent in a population retains a meaningful version.
6. Illustrate with small hand-computable examples. Nearly every concept in the paper is demonstrated with 2- or 3-dimensional vectors and diagonal or small rectangular matrices so readers can verify the arithmetic. Valuation is simplified to a vector norm or a sum of components.
7. State the main results as theorems and push details to appendices. Appendix B.1 covers the formalization of value spaces and motivation, B.2 abstract beings, B.3 consistency and interpretation maps, B.3.1 the local coherence condition and an informal piecewise-linear extension of the Local Coherence Theorem, B.4 null spaces, B.5 networks, and B.11 the lifecycle (birth, evolution, death) of abstract beings.
8. Preview an empirical check. Section 4.3 is described as an empirical illustration that uses sentence-embedding spaces to estimate T between philosophical and engineering vocabularies. The full content of that section, and its results, are not reported in the available text.

Why This Matters

Impact on research. The paper offers a way to talk about misunderstanding, value misalignment, motivational drift, and cultural innovation with one shared algebraic vocabulary instead of separate accounts for each. It reframes disagreement between ideally rational agents as a structural property of their representational geometries, which could change how AI alignment, social epistemology, and conceptual-space research frame their problems. It also supplies an explicit test for when alignment is structurally impossible rather than merely incomplete.

Real-world applications the framework addresses or implies:

• Cross-disciplinary and cross-cultural communication. The shared-basis hypothesis predicts that concepts will be largely annihilated when they cross the boundaries of a shared task subspace — engineering versus philosophy, clinical versus lay reasoning — because the receiver lacks the basis needed to recover them.
• AI value alignment and interpretability. The null-space criterion gives a structural test for when an AI system's internal representation of a goal diverges from a human's intended meaning, complementing the normative and empirical approaches of Gabriel (2020) and Russell (2019).
• Leadership and organizational influence. The No-Null-Space Leadership Condition treats leadership as representational reachability: an agent can lead on a dimension only if that dimension is not annihilated in followers' spaces. Innovation additionally requires occupying a vector outside the convex hull of the group's current valuations.
• Marketing and cultural change. The paper states that the model explains marketing-induced reconfiguration of value spaces, and that a being's extinction across a population corresponds to concept death, with collective memory represented as the cognitive space spanned by all agents' conceptual structures.

Industry relevance. Any setting where an organization or a product must move meaning between people with different internal value systems — team coordination, product messaging, human-AI interaction, safety-critical AI systems whose goals are specified by humans — faces exactly the transformation problem this framework formalizes. The framework's practical value is diagnostic: it locates the failure in the geometry of the receiver rather than in the sender's clarity or the receiver's reasoning.

Future Directions

• Nonlinear interpretation maps. The paper states that a fully nonlinear treatment remains future work; only an informal piecewise-linear extension of the Local Coherence Theorem is given in Appendix B.3.1.
• General valuation functions. Theorems 5 and 6 in the appendix assume Val_i is a norm or a component sum (Remark 1), and the paper explicitly names extending them to general valuation functions as a direction for future work.
• Field-level basis-sharing dynamics. A full treatment of how shared bases form, fracture, and are renegotiated across communities is described as a planned extension to epistemic flow across disciplines.
• Empirical estimation of interpretation maps. Section 4.3 is described as estimating T between philosophical and engineering vocabularies using sentence-embedding spaces; scaling this kind of estimation and validating the framework's predictions against observable communication remain open.
• Unresolved theoretical question: the paper poses, but does not fully answer with data, the question of which empirical regularities make interpretation maps approximately invertible on the concepts that matter.

Target Audience

Researchers and graduate students in AI theory, AI value alignment and interpretability, cognitive science and conceptual-space modeling, formal epistemology, philosophy of mind and language, multi-agent systems, and cultural evolution. The paper is also relevant to readers interested in leadership, influence, and collective memory who are comfortable with basic linear algebra, since the formal machinery is elementary even where the philosophical framing is ambitious. Readers looking for empirical benchmarks, datasets, or model comparisons will not find them here — the paper reports no datasets, no benchmark evaluations, and no quantitative experiments beyond its own hand-computed illustrative examples.

Authors’ abstract

This paper develops a geometric framework for modeling concepts, motivation, and influence across cognitively heterogeneous agents. Each agent is represented by a personalized value space, a vector space encoding the internal dimensions through which the agent interprets and evaluates meaning. Evaluative concepts are formalized as structured vectors, abstract beings, whose transmission is mediated by linear interpretation maps. An abstract being survives communication only if it avoids the null spaces of these maps, yielding a structural criterion for intelligibility, miscommunication, and concept death. Within this framework, I show how conceptual distortion, motivational drift, and the limits of mutual understanding arise from purely algebraic constraints. A central result, the No-Null-Space Leadership Condition, characterizes leadership as a property of representational reachability rather than persuasion or authority. More broadly, the model explains how abstract beings can propagate, mutate, or disappear as they traverse diverse cognitive geometries. The account unifies insights from conceptual spaces, social epistemology, and AI value alignment by grounding meaning preservation in structural compatibility rather than shared information or rationality. I argue that this cognitive-geometric perspective clarifies the epistemic boundaries of influence in both human and artificial systems, and offers a general foundation for analyzing conceptual dynamics across heterogeneous agents.