A new paper identifies a significant documentation gap in medical literature, termed LABEL (Literature Asymmetry between Efficacy and Limits). This asymmetry shows that drug benefits are mentioned about twice as often as safety contraindications across sources like Wikipedia and PubMed. Language models, including Qwen2.5, mirror this gap, performing poorly on contraindication accuracy while improving on benefit accuracy. However, models demonstrate capability when provided with retrieved passages or when safety questions are routed to official FDA labels, suggesting the issue stems from the scarcity of safety information in general text corpora. AI
IMPACT Highlights the need for AI systems to prioritize official regulatory data for medication safety, rather than relying on general text corpora.
RANK_REASON Research paper detailing a novel finding about information asymmetry in medical literature and its impact on language models. [lever_c_demoted from research: ic=1 ai=1.0]
Read on arXiv cs.IR (Information Retrieval) →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →