Title: Automated Detection of Lexical Alignment and Preference-Stage Shifts in Large Language Models

Abstract: The linguistic output of digital chat assistants, including ChatGPT, often deviates from human expectations, a phenomenon known as misalignment. While previous research, largely confined to Scientific English, has documented these divergences and partially explained their causes by linking them to the human preference learning stage, existing methods have depended on manual curation. This study presents two novel evaluation metrics that require no manual intervention and rely on minimal assumptions: the Lexical Alignment Score, designed to detect lexical overuse, and the Triangulated Preference Shift, which quantifies the extent to which such shifts result from human preference learning. We generated continuations using PubMed abstracts and evaluated them across six model families (Falcon, Gemma, Llama, Mistral, OLMo, and Yi) by measuring windowed document prevalence. Our automated approach successfully identified overused terms such as 'suggest', 'additionally', and 'strategy', and estimated their correlation with preference learning processes. The results align with prior findings and demonstrate robustness across varying parameter settings, random seeds, and additional datasets. This scalable method facilitates the systematic investigation of lexical (mis)alignment beyond the realm of Scientific English and across multiple languages, offering potential insights for enhancing model alignment and understanding its underlying mechanisms.

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