Why 3A-LLM qualifies as a large language model

3A-LLM is not a neural transformer trained on corpora; we call it a “large language model” because it models the conceptual space underlying language and exhibits large generative scope. First, 3A-LLM focuses on semantics: it models the conceptual space rather than the lexical surface, aligning with the view that languages serve communication and concepts are at their core. Second, through the transformation calculus and compounding, the model constructs arbitrarily deep compound structures, each with full transparency and reversibility. Third, the graph-theoretic organisation spans many semantic fields and supports cross-domain inference [Gardenfors2000]. Fourth, 3A-LLM is inherently cross-lingual: the conceptual graph is language-independent; lexical items from any language map to the same concept node. Each concept node can carry language-tagged surface forms (e.g. table with table.en=“table”, table.de=“Tisch”, table.es=“mesa”). A query with a surface form in any language is resolved to the underlying concept via reverse lookup; from that point, inference and expansion operate on the same conceptual structure, and results can be returned in any target language. Query and inference thus work identically across languages without retraining or statistical alignment. Fifth, Cantor-based encoding operationalises this conceptual largeness in storage and indexing. Table ? summarises the comparison; 3A-LLM's “largeness” derives from semantic breadth, generative depth, and structural richness [Zhao2023].

Although the Axiomatic Large Language Model (A-LLM) does not rely on billions of neural parameters, it nevertheless satisfies the defining criteria of a Large Language Model in a conceptual, structural, and generative sense. The notion of “large” in LLM research has historically been tied to parameter count only because neural architectures required massive weight matrices to approximate linguistic structure [Zhao2023], [Minaee2024]. A-LLM demonstrates that largeness can be achieved symbolically rather than statistically, and that linguistic scalability need not be tied to numerical weight magnitude.

We take a Large Language Model to be a computational system that (a) represents or processes natural language, and (b) possesses large-scale expressive, structural, or statistical capacity in at least one of the following dimensions: a large parameter space (e.g., neural weights), a large conceptual space (e.g., semantic graphs, definitional primitives), a large generative space (systematic derivation of arbitrarily many expressions), or large coverage of language (domains, constructions, cross-lingual mappings). Thus, “large” does not require neural parameters; a model is “large” if it scales meaning, structure, or expressiveness to a degree comparable to or exceeding neural LLMs.

First, A-LLM possesses a large conceptual space. Beginning from approximately 1500--2000 LDV primitives, the model generates an unbounded number of derived concepts through vertical transformations, horizontal geometric operators, and functional compounds. The resulting semantic graph contains a theoretically infinite number of nodes, each definitorially traceable, making the model “large” in the sense of conceptual coverage rather than parameter count.

Second, A-LLM exhibits a large generative space. Through the algebraic calculus described in Sections 4--6, the model can construct arbitrarily deep compound structures---such as instrument(pressure(air)) or verb(opposite(largeness))---each representing a unique conceptual object. This level of recursive generativity mirrors or exceeds the combinatorial capacity of neural LLMs, but with complete transparency and reversibility.

Third, the model's structural largeness arises from its graph-theoretic organization. The directed and typed semantic graph spans numerous semantic fields, supports complex inferential paths, and enables cross-domain modeling [Gardenfors2000]. In contrast to neural vector spaces, where relational structure is implicit, A-LLM encodes semantic largeness explicitly through topological and algebraic relations among concepts.

Fourth, A-LLM supports large-scale cross-lingual coverage. By mapping natural-language surface forms onto language-agnostic conceptual nodes, the model abstracts away from linguistic variety and enables multilingual reasoning without corpus alignment. The conceptual substrate remains constant even as lexical forms vary across languages, giving A-LLM a truly large semantic domain.

Finally, Cantor-based integer encoding ensures that this conceptual largeness can be operationalized in computational environments. Arbitrarily complex compounds can be encoded as single integers, enabling efficient storage and transmission while preserving full decomposability. This numeric scalability reinforces the model's classification as “large” by providing an unbounded and reversible representation of conceptual structure.

Taken together, these properties justify the classification of A-LLM as a Large Language Model. Its largeness derives not from statistical parameters but from semantic breadth, generative depth, structural richness, and cross-lingual universality---dimensions of scale that are at least as significant as the weight magnitude that defines neural LLMs [Zhao2023].

Table ? summarizes how A-LLM satisfies the criteria of a Large Language Model in comparison to classical neural LLMs. The table is presented in landscape orientation to fit the page.