Scaling Scientific Discovery Environments for Turn-Level Agentic RL
Yucheng Xu, Keyi Zhang, Yuyang Yu, Min Zhang, Shiyuan Meng, Pei Chu, Zhongying Tu
- Digest date
- 2026-08-03
- Submitted
- 2026-07-31
- arXiv ID
- 2607.28990
Summary
SciDisco builds process-verifiable environments (SciThèque) compiling hypotheses, datasets, hidden evidence graphs, and verifiers, then uses DAG-grounded trajectory synthesis and DiscoPO to assign turn-level credit for verifiable analytical progress during multi-turn scientific data analysis. SciDisco-14B reaches state-of-the-art on hypothesis-driven scientific analysis benchmarks.
Abstract
Large language model agents have shown promising capabilities in data-driven scientific discovery tasks, where an agent interacts with an execution environment and produces a statistical claim. Long-horizon scientific analysis remains constrained by the lack of process supervised environments over real-world scientific data. This paper introduces SciDisco, a scalable framework for training Scientific Discovery agents in process-verifiable environments. SciTh\`eque compiles hypotheses, datasets, hidden evidence graphs, and verifiers into task environments where analytical progress can be checked during interaction. DAG-grounded trajectory synthesis uses these environments to construct verifier-filtered multi-turn demonstrations. DiscoPO then uses the environment as the source of training signal, assigning turn-level credit to actions that produce verifiable analytical evidence. Experiments show that SciDisco-14B reaches state-of-the-art on hypothesis-driven scientific data analysis benchmarks.