Isolating LLM Lexical Bias: A Curation-Free Triangulated Metric for Preference-Stage Learning

In three linesNew automated metric (Triangulated Preference Shift score) to measure lexical biases introduced during preference learning (RLHF) in LLMs without manual curation. Analysis across 6 model families reveals a shift toward a 'language of prestige' (overuse of 'delve', 'furthermore').

Summary generated by Claude — human-verified