arXiv:2606.02750v2 Announce Type: replace 
Abstract: Representations extracted from large language models (LLMs) play an important role in many downstream applications. However, the structure of these representations is often influenced by lexical overlap rather than semantic content. Our understanding of the relationship between this lexical influence and semantic content, and its implications for downstream tasks, remains limited. In this work, we investigate representations to quantify the effect of lexical overlap relative to semantic content. We consider several adversarial semantic stress tests and further connect our findings to the information theory perspective. We find that lexical influence extends across the depth of models, consistently across architectures, training regimes, and objective functions, including the models trained for semantic similarity. Moreover, we observe a mid-depth region in which both lexical and semantic signals degrade simultaneously, indicating a transitional regime where representations are poor for both surface form and meaning. We further demonstrate the effect of lexical influence on downstream uses of LLMs using summarization and model editing as a case study.

دراسة حديثة نُشرت على arXiv تبحث في الآثار المستمرة لل lexicality في نماذج اللغة الكبيرة (LLMs)، كاشفةً أن التداخل اللفظي يؤثر بشكل كبير على هيكل التمثيلات المستخرجة من هذه النماذج، غالبًا ما يتجاوز المحتوى الدلالي. تستخدم الدراسة اختبارات ضغط دلالي عدائية لتحديد هذه التأثيرات عبر هياكل مختلفة وأنظمة تدريب.

Un estudio reciente publicado en arXiv investiga los efectos persistentes de la lexicalidad en los modelos de lenguaje de gran tamaño (LLMs), revelando que la superposición léxica influye significativamente en la estructura de las representaciones extraídas de estos modelos, a menudo eclipsando el contenido semántico. La investigación emplea pruebas de estrés semántico adversarias para cuantificar esta influencia a través de diversas arquitecturas y regímenes de entrenamiento.

Une étude récente publiée sur arXiv examine les effets persistants de la lexicalité dans les modèles de langage de grande taille (LLMs), révélant que le chevauchement lexical influence de manière significative la structure des représentations extraites de ces modèles, souvent au détriment du contenu sémantique. La recherche utilise des tests de stress sémantique adversariaux pour quantifier cette influence à travers diverses architectures et régimes d'entraînement.

A recent study published on arXiv investigates the persistent effects of lexicality in large language models (LLMs), revealing that lexical overlap significantly influences the structure of representations extracted from these models, often overshadowing semantic content. The research employs adversarial semantic stress tests to quantify this influence across various architectures and training regimes.