New bias 'Narrative Anchoring' found in clinical LLMs

Researchers have identified a new bias in clinical language models called Narrative Anchoring, where identical medical facts presented in different linguistic styles can lead to divergent diagnoses. This bias was observed across multiple large language models, even when demographic information was absent. To address this, a method called NarrativeShield was developed, which extracts and verifies clinical facts before diagnosis, significantly reducing the bias and improving decision stability with a minor impact on accuracy. AI

IMPACT Highlights a new vulnerability in LLMs for critical applications like healthcare, necessitating robust bias detection and mitigation strategies.

RANK_REASON Academic paper detailing a new bias in LLMs and proposing a mitigation method. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Hugging Face
• Narrative Anchoring
• Narrative Anchoring Gap
• NarrativeShield
• United States Medical Licensing Examination

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →