SMDD-Bench tests whether LLM agents can budget through real small-molecule design

niloofar_mire · x · 2026-07-23

Carnegie Mellon researchers say their AI-for-science work, including SMDD-Bench, is one of three projects selected for U.S. Department of Energy Genesis Mission awards.

The referenced benchmark, SMDD-Bench, evaluates long-horizon agentic small-molecule design. It includes 502 guaranteed-solvable tasks across five real drug-design workflows — pharmacophore identification, scaffold hopping, lead optimization, fragment assembly, and interaction point discovery — and gives agents a Python sandbox plus strict limits on Boltz2 and ADMET-AI calls. The goal is to test whether frontier LLM agents can plan and budget through multi-turn medicinal chemistry tasks instead of handling toy single-step prompts.

Related event: CMU Receives Three DOE Awards for AI-Driven Molecular Design(2 posts)→

Original post →

More from Companies & People

Companies & People channel →