Study: LLMs rely on morphological shortcuts in drug names

Arxiv CS.CL (NLP/language models)

LLMs exploit morphological cues in drug names to reason about fictitious compounds, indicating overgeneralization in high-stakes pharmacology contexts. The study highlights risks of relying on word-form mappings.

llm

Arxiv CS.CL (NLP/language models)·Kaijie MoThomas YangChantal ShaibQing YaoWilliam RudmanRamez KouzyKanishka MisraByron C. WallaceJunyi Jessy Li

6 days ago