Large Language Models as Quasi-crystals: Coherence Without Repetition in Generative Text

TL;DR

Problem: LLM coherence is typically assessed in semantic or factual terms, potentially overlooking long-range structure. Approach: introduce a quasicrystal-inspired analogy that treats LLM outputs as constraint-driven, non-repeating forms and propose a structural evaluation framework. Contributions: articulate the concept of structural coherence, outline methods like entropy mapping and motif analysis, and connect the approach to structural realism and inferentialism. Significance: offers a new interpretive axis for designing, evaluating, and trusting generative systems where form and pattern matter as much as content.

Abstract

This essay proposes an interpretive analogy between large language models (LLMs) and quasicrystals, systems that exhibit global coherence without periodic repetition, generated through local constraints. While LLMs are typically evaluated in terms of predictive accuracy, factuality, or alignment, this structural perspective suggests that one of their most characteristic behaviors is the production of internally resonant linguistic patterns. Drawing on the history of quasicrystals, which forced a redefinition of structural order in physical systems, the analogy highlights an alternative mode of coherence in generative language: constraint-based organization without repetition or symbolic intent. Rather than viewing LLMs as imperfect agents or stochastic approximators, we suggest understanding them as generators of quasi-structured outputs. This framing complements existing evaluation paradigms by foregrounding formal coherence and pattern as interpretable features of model behavior. While the analogy has limits, it offers a conceptual tool for exploring how coherence might arise and be assessed in systems where meaning is emergent, partial, or inaccessible. In support of this perspective, we draw on philosophy of science and language, including model-based accounts of scientific representation, structural realism, and inferentialist views of meaning. We further propose the notion of structural evaluation: a mode of assessment that examines how well outputs propagate constraint, variation, and order across spans of generated text. This essay aims to reframe the current discussion around large language models, not by rejecting existing methods, but by suggesting an additional axis of interpretation grounded in structure rather than semantics.