CIPHER Improves Data Science Agents with Decoupled Exploration-Selection.

Automating complex data science tasks, from information extraction to open-ended analysis, presents significant challenges. While recent AI agents powered by large language models show promise, their reliance on a single initial state can lead to cascading errors if that state is suboptimal. This research introduces CIPHER, an innovative automated data science agent designed to overcome this limitation through "test-time scaling." CIPHER's core innovation is its Decoupled Exploration-Selection (DES) framework. Unlike existing approaches, CIPHER explicitly separates the process of generating multiple candidate initial states from the strategic selection of which states to execute in parallel. This decoupling allows the agent to explore a wider range of potential starting points and choose the most promising ones, significantly mitigating the risk of early-stage errors. Extensive evaluations on both closed-form and open-form data science tasks demonstrate that CIPHER achieves state-of-the-art performance, even outperforming larger-model baselines while using a substantially smaller base language model. The study also provides actionable design recommendations for practitioners by characterizing how generation strategy, selection strategy, and aggregator model capacity influence overall performance.