Translucent: Language agents for autonomous experimental engineering

Abstract

Models, simulators, automated instruments and agents increasingly support science, yet much work still lies between acquiring experimental evidence and deciding what to test next: interpreting instrument records, establishing what measurements mean, choosing and checking models and defining the experiment. We ask whether a language agent can take responsibility for that preparation and carry an investigation forward through experimental feedback. Translucent runs a general language model, Muse Wonder, that proposes every import, analysis, fit, check, measurement and design request; numerical tools compute, a server enforces approvals and an evidence rule and keeps a persistent record, and a person approves any physical test. In one battery investigation the model turned two unfamiliar cycler exports into a supported fast-charging design after finding that the source cells disagreed and choosing simulated measurements of the target fixture. From raw exports, Translucent recommended in 5/6 runs, all supported and feasible in hidden simulation; the same model in a plain tool loop recommended in 6/6, one unsupported. With eight simulated experiments, Translucent's median charge times were 33.79, 37.17 and 34.13 min versus 42.62, 42.62 and 45.98 min for constrained Bayesian optimization, and every Translucent run beat every run of the same model without experimental feedback, which makes fewer model calls. A prepared numerical procedure matched its model-learning designs with fewer tests. All charging experiments were simulated.

Code and records

• Implementation and evaluation reports: github.com/ychampion/translucent
• The investigation the paper walks through, replayable in the browser: Raw records to a design