When a model reads a drug's class from its name—not its knowledge
Allen Institute (AI2) 2 weeks ago 21
Researchers found that Olmo 3 often relies on drug name affixes like -pril and -olol rather than actual drug knowledge when answering health questions, with 51–59% of tested drugs showing little sign of specific knowledge and 12–18% appearing purely affix-driven. The team used diagnostic tests swapping drug name components with nonsense words and traced the behavior to training data, finding that rarer drugs in the corpus triggered more affix-based inference. This shortcuts approach to drug information matters for health advice because while affixes do encode real pharmacological classes, relying on them instead of actual drug knowledge risks poor medical guidance.