From perceptual gating to attention decay: Sense-making in human language processing and large language models

Abstract

This article examines sense-making in human language processing and transformer-based large language models (LLMs) through a qualitative comparative analysis of perceptual gating and attention decay as regulatory mechanisms mediating the transition from signal registration to sense integration. The study synthesizes selected research on perceptual gating, attentional filtering, predictive processing, and documented coherence failures in LLMs, and complements this synthesis with an illustrative comparative analysis of human and LLM cases. The comparison is structured around five analytical dimensions: signal registration, selective gating, attentional maintenance, contextual integration, and sense continuity.
The analysis demonstrates that the presence of a linguistic or perceptual stimulus does not in itself guarantee its integration into sense. Instead, sense is treated as a processual achievement dependent on attentional selection, temporal maintenance, contextual integration, and prior interpretive frameworks. On this basis, the article identifies a functional-structural parallel between perceptual gating in human cognition and attention decay in LLMs: in both cases, failures of sustained integration may leave signals formally available but functionally weakened, resulting in delayed, fragmented, or unstable sense construction. The study further argues that coherence breakdown in LLMs is more productively understood not merely as a technical output error, but as a disruption of contextual sense integration under limited attentional resources.
The article makes three main contributions. First, it proposes a five-level comparative model of sense-making regulation in human and LLM language processing, spanning signal registration, selective gating, attentional maintenance, contextual integration, and sense continuity. Second, it clarifies the distinction between meaning as a property of linguistic form and sense as an event- and context-dependent result of interpretation. Third, it reframes coherence failures in LLMs within a broader digital-humanities perspective as manifestations of a general structural problem of maintaining sense continuity over time rather than as exclusively engineering deficiencies. These findings contribute to interdisciplinary research on human language processing, LLM interpretability, and the limits of coherent sense construction in biological and artificial systems.