UT Austin study: LLMs infer drug properties from name suffixes, hallucinating fake drugs
jessyjli · x · 2026-10-16
New research from UT Austin systematically quantifies morphology-driven inference in LLMs across 653 drugs: models often guess drug properties from affix cues like "-cillin" rather than real pharmacological knowledge. Using paired real (ampicillin) vs fictitious (dimicillin) drug questions, the study shows models confidently attribute antibiotic properties to fake drugs, rarely disclose this reliance, and sometimes conflate properties among affix-sharing drugs. The behavior is analyzed at both behavioral and mechanistic levels.
!
More from Research
- Cornell's Copy-Paste LMs add a lightweight head to boost in-context copying and long-context retrieval — lil-lab · 2026-10-16
- Study explains expert roles in MoE LLMs, enabling localized and interpretable personalization — copenlu · 2026-10-16
- SCALLOP: NeurIPS 2026 paper distills likelihood flow maps for few-step Boltzmann generation — jmhernandez233 · 2026-10-16
- Terence Tao warns AI is clearing math's open problems faster than it creates new ones — VoidStateKate · 2026-10-16
- TMLR suspends 159 authors, lifts desk rejections to 53% amid 300% submission surge — thegautamkamath · 2026-10-16
- TabPFN sweeps Kaggle: now in 17% of public notebooks and 3 of September's top-10 solutions — ph_singer · 2026-10-16