Operational limits of ontological innovation in large language models: evidence from controlled sensory-modality generation
Type: article
Abstract: Contemporary large language models generate text that can appear conceptually novel, raising the question of whether outputs reflect genuine ontological innovation-the introduction of concepts that cannot be expressed as recombinations or transformations of training-derived primitives-or sophisticated recombination within a fixed representational manifold.We address this question through a controlled empirical study using an eight-modality multi-modal sensory ontology as the experimental universe.Seven state-of-the-art language models (N = 168 proposals, 24 runs each) were prompted to propose a ninth sensory modality genuinely beyond the defined eight.Outputs were evaluated through four convergent methods: literature traceability with calibrated threshold (τ = 0.40), structural decomposition, embeddingspace convex-hull analysis, and continuous novelty scoring.Results show a strict novelty rate of p novel = 0.000 (95% CI: [0.000, 0.000]), a traceability rate of p traceable = 0.982, and 100% convex-hull membership under Sentence-BERT embeddings reduced to three principal components.Structural decomposition revealed three failure modes: range extension (20.2%), hybrid combination (44.6%), and functional recombination (17.9%).Cross-model effects were moderate (η 2 = 0.326), but all models stayed within the same low-novelty regime, supporting a constrained-manifold rather than a model-deficit explanation.We interpret these findings through the No New Basis Theorem: gradient-descent learning over fixed vocabularies cannot expand representational dimensionality beyond training data.These results support the view that current neural architectures are bounded within training-derived conceptual closures, with implications for operationalizing and evaluating ontological creativity.
Source: CyberLeninK (CyberLeninka)
Author: Matarmaa Jarno Olavi
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223542206