Raven: From Scientific Data to Trustworthy Autonomous Discovery

Abstract

Transformational AI for science will depend less on isolated model capability than on the quality of the scientific data interfaces surrounding each model. Raven is a state-aware agentic architecture for turning heterogeneous data, simulations, literature, and experiments into auditable discovery decisions. Its central design choice is to represent each step as a typed, provenance-carrying state transition and to expose local models and instruments as verifiable capabilities. A materials-discovery testbed demonstrates how domain-specific data, forward-model verification, and reinforcement learning from verifiable rewards can connect scientific data infrastructure to experimentally testable action. Raven: \url{https://ravenllm.com/}