Study: LLMs Fake Drug Knowledge by Relying on Naming Patterns
allen_ai · x · 2026-08-19
Researchers found that LLMs often lack specific drug knowledge and rely on morphological clues in names (e.g., "-pril" for ACE inhibitors) instead. By using the open OLMo model and the infini-gram tool, they traced the training data to confirm this behavior.
Related event: LLMs fake drug knowledge by guessing from name suffixes, study finds(2 posts)→
More from Research
- Qwen 3.8 Comparison: Smaller Q4 Model Outreasons Larger Q5 — k-r-a-u-s-f-a-d-r · 2026-08-19
- Open Source: Acoustic UAV Detection for Battlefield Scenarios — yehors · 2026-08-19
- ClawGym II paper: Improving agents via mixed-harness training — omarsar0 · 2026-08-19
- Rich Sutton: Synthetic Data Is a Mistake, LLMs Are Only a Quarter of Intelligence — GregCook2011 · 2026-08-19
- Study: Context compactor hides real costs as Agent retrieval calls triple — dair_ai · 2026-08-19
- Paper: LLMs severely underestimate missing information, study finds — marinkazitnik · 2026-08-19