CMU team releases SMDD-Bench: 502 small-molecule drug design tasks for RL agents
willcb · x · 2026-10-03
A year after pivoting to AI for molecule design and drug discovery, Niloofar Gheini's CMU team, together with PrimeIntellect, released SMDD-Bench and its training environments.
- Scale: 502 small-molecule design tasks with RDKit, ADMET-AI, and Boltz-2 in the loop, available on PrimeIntellect's Environments Hub, ready to train with prime-rl.
- Motivation: RL needs long-horizon tasks beyond math and coding; drug design adds open problems in long-horizon planning, exploration, and learning from imperfect feedback.
- Findings: 1) Harness optimization goes a long way in scientific tasks, and progress is still bottlenecked by model capabilities rather than knowledge; 2) Automating harness optimization works in some cases but not others.
A notable effort to extend RL/agent evaluation into the sciences.
More from Research
- LEV: Looped decision model lifts 2.6B accuracy 63.2% to 76.2% with adaptive compute — jiqizhixin · 2026-11-19
- Aurora2-400M-A20M: a sparse MoE pretrained from scratch in 36 hours on one RTX Pro 6000 — Tall_Abrocoma_3533 · 2026-11-19
- Researcher calls METR/Anthropic '2x speedup' RSI analysis statistically dubious — jessi_cata · 2026-11-19
- Linear Probes Could Reveal Which Persona a Base Model Is Playing, Researchers Suggest — cephaloform · 2026-11-19
- ETH's Anymal robot completes 10km autonomous alpine hike with VLM-guided ROAM system — ChongZzZhang · 2026-11-19
- OpenVax codebase grows from 78K to 565K lines in 15 months — iskander · 2026-11-19