A new arXiv paper investigates how post-training techniques affect Large Language Model (LLM) fact-checking behavior, specifically focusing on 'job-based doubt,' where models may judge claims as false more often if posted by individuals with lower-status jobs. The research found that only the open Tülu-3 recipe, utilizing OLMo-2 32B and Tülu-3 70B models, introduced this subtle but statistically reliable bias. Interestingly, the models did not necessarily pay more attention to the poster's job but rather interpreted lower-status jobs differently, shifting from a 'true' inclination in the base model to a 'false' inclination after post-training. This shift was traced to a single residual-stream direction during supervised fine-tuning, a phenomenon not observed in commercial models from Meta, Alibaba, or Mistral AI. AI
IMPACT Highlights the need to audit speaker bias introduced by specific LLM training recipes, potentially impacting how AI systems evaluate information based on source.
RANK_REASON Academic paper detailing a specific finding about LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
