This research evaluates how language models handle contradictions between their internal knowledge and external evidence in retrieval-augmented generation (RAG) systems.

• Introduces ClashEval, a framework to quantify the tug-of-war between LLM priors and retrieved information
• Analyzes how different models react to incorrect information in RAG contexts
• Reveals variation in how models handle evidence across domains like medicine and education
• Provides insights for developing more robust RAG systems that appropriately weigh evidence

Medical Impact: Critical for medical AI applications where models must properly evaluate drug dosage information to prevent harmful outcomes from incorrect retrieved content.

ClashEval: Quantifying the tug-of-war between an LLM's internal prior and external evidence