Fully Automated Identification of Lexical Alignment and Preference-Stage Shifts in Large Language Models

In three linesTwo automated metrics assess LLM lexical misalignment: Lexical Alignment Score detects term overuse ('suggest', 'additionally', 'strategy'), Triangulated Preference Shift quantifies RLHF impact. Tested on 6 model families (Falcon, Gemma, Llama, Mistral, OLMo, Yi) via PubMed abstracts, no manual annotation required.

Summary generated by Claude — human-verified