New Benchmark Reveals LLMs Struggle With Real-World Drug Design Tasks, Solving Only 40% of Cases

Researchers introduced SMDD-Bench, a standardized benchmark with 502 drug design tasks spanning multiple chemistry types and protein targets, to evaluate how well large language models can handle autonomous molecular discovery. Testing seven frontier LLMs showed that even the best performer, GPT-4, solved only 40.2% of tasks, suggesting significant gaps remain in LLM reasoning for complex chemical and biological problems.

Why it matters: As AI investment in drug discovery accelerates, this benchmark establishes the first rigorous, multi-turn evaluation standard for LLM agents in pharma—critical for separating genuine breakthroughs from marketing claims.