Curie: Toward Rigorous and Automated Scientific Experimentation with AI Agents
What this work does
Architect/technician agents interact through modules that validate experimental procedures, partition plans, enforce dependencies and retain structured experiment knowledge.
Connection to discovery
A concrete harness contribution: experimental control and provenance can be evaluated independently of policy parameter updates.
Evidence
Inspected validation and scheduling rules, the 46-question four-domain benchmark, and conclusion-accuracy results.
Scope and limitations
Computer-science experiments dominate; strong within-benchmark gains do not show transfer to physical discovery. Validation modules themselves can miss confounds.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.