SciDisco: A Scalable Framework for Turn-Level Agentic RL in Scientific Discovery

A recent study published on arXiv (2607.28990) presents SciDisco, a scalable framework designed for training agents focused on scientific discovery within process-verifiable settings. This framework tackles the complexities of long-term scientific analysis by offering process supervision based on actual scientific data. SciDisco consists of three main elements: SciThèque, which organizes hypotheses, datasets, hidden evidence graphs, and verifiers into interactive task environments; DAG-grounded trajectory synthesis, which utilizes these environments to create verifier-filtered multi-turn demonstrations; and DiscoPO, which leverages the environment to provide training signals, crediting actions that yield verifiable analytical evidence. Experiments indicate that SciDisco-14B achieves state-of-the-art results on benchmarks for scientific discovery. The authors of the paper are researchers, and it can be accessed on arXiv.