SciDisco Framework Advances AI Agents for Scientific Discovery

The field of scientific discovery is increasingly leveraging large language model agents, but their potential for long-horizon analysis has been hampered by a lack of process-supervised environments using real-world scientific data. To overcome this, researchers have introduced SciDisco, a scalable framework designed for training Scientific Discovery agents in environments where analytical progress can be verified at each step. SciDisco incorporates SciThèque, a component that compiles hypotheses, datasets, hidden evidence graphs, and verifiers into comprehensive task environments. These environments are then used to construct verifier-filtered, multi-turn demonstrations through DAG-grounded trajectory synthesis. This ensures that the training data is grounded in verifiable scientific processes. A key innovation within SciDisco is DiscoPO, which utilizes the environment to provide turn-level training signals. This module assigns credit to agent actions that produce verifiable analytical evidence, allowing for more precise and effective reinforcement learning. Experiments demonstrate that SciDisco-14B achieves state-of-the-art performance on benchmarks for hypothesis-driven scientific data analysis, marking a significant step forward in autonomous scientific research.