Autonomous Science Bench: Evaluating AI Agents for Autonomous Scientific Discovery
What this work does
Runs budgeted design-build-test-learn campaigns, judging both the quality of measured designs and prediction/ranking of held-out designs.
Connection to discovery
A directly relevant benchmark for the separability of finding promising experiments and learning a reusable model under a limited measurement budget.
Evidence
The paper specifies five recorded-data replay campaigns and one digital-twin campaign, with a live physical backend described as planned. Gate calibration and failure decompositions distinguish acquisition success from predictive learning.
Scope and limitations
The initial campaigns use predefined design spaces and existing measurements or a surrogate. A lab-service API alone is not evidence of new physical execution; calibrated gate choices also shape the measured capability.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.