Title: The Enduring Impact of Lexicality in Large Language Models

Abstract: Extracted representations from large language models (LLMs) are critical for numerous downstream applications. Yet, the architecture of these representations is frequently shaped by lexical overlap instead of semantic substance. Our comprehension of how lexical influence interacts with semantic content, and what this means for practical applications, remains constrained. This study examines LLM representations to measure the magnitude of lexical overlap in comparison to semantic content. We employ various adversarial semantic stress tests and relate our results to an information-theoretic framework. Our analysis reveals that lexical influence permeates the entire depth of the models, a pattern that persists across different architectures, training methods, and objective functions, even among models specifically optimized for semantic similarity. Additionally, we identify a transitional zone in the middle layers where both lexical and semantic signals deteriorate concurrently, suggesting a phase in which representations are ineffective for capturing both surface-level forms and deeper meaning. To illustrate the real-world consequences of this lexical influence, we analyze its impact on downstream LLM utility through case studies involving summarization and model editing.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC