A new preprint investigates how Large Language Models (LLMs) evaluate arguments based on their attributed source. The study found that credibility was the most sensitive dimension to source attribution, often showing an "irrelevant-expertise penalty" where attributing an argument to an expert in an unrelated field reduced its credibility compared to having no source. The models tested included GPT-5.6 Sol, Claude Opus 5.5, Gemini 4 Argon, and DeepSeek Flash, evaluated on logic, credibility, persuasiveness, and agreement. AI
IMPACT Reveals potential biases in LLMs related to source attribution, impacting how their evaluations should be interpreted.
RANK_REASON The cluster reports on a new preprint detailing research into LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
