Benefit Ledger · verified · Chemistry
Carnegie Mellon researchers showed a GPT-4-driven agent, Coscientist, that searched literature, wrote robot code, and optimized palladium-catalyzed couplings, published in Nature on 20 December 2023.
Current score
+0.26
Notable automation result (tier 2). GPT-4 was necessary but humans built tools and hardware. Peer-reviewed demo. Realization is lab tasks, not new molecules in the clinic. Medium durability as agent stacks age quickly.
Coscientist combined GPT-4 with tools for web search, documentation, code execution, and lab APIs, including an Opentrons OT-2. Demonstrations included Suzuki–Miyaura and Sonogashira planning and reaction optimization with GC-MS confirmation. Claude was also used in some configurations. The work is a controlled lab demonstration, not a general autonomous scientist.
Primary planner LLM that called tools and wrote experimental code.
Paper describes a GPT-4-driven multi-LLM agent; humans supplied tools and robots.
Attribution 0.6000 · Credit share 70% · OpenAI
Coscientist planned and executed palladium-catalyzed coupling experiments via robotic APIs.
outcome · supported
GPT-4 was the primary language model in the published system.
attribution · supported
primary sources
Nature · peer reviewed · 20 Dec 2023 · peer reviewed
Carnegie Mellon University · university announcement · 20 Dec 2023
Secondary domains: Computer Science
13 Sep 2026 · 0.00 → 0.26
Initial adjudicated seed score under methodology 0.1.