CIPHER: A Decoupled Exploration-Selection Framework for Test-Time Scaling of Data Science Agents

arXiv cs.AI · 6 hours ago

Researchers introduced CIPHER, a data science agent framework that generates and selects multiple initial states for concurrent execution to improve performance on complex tasks. The system was evaluated on two benchmarks and achieved state-of-the-art results in matched-model comparisons while using a substantially smaller base language model than larger-model baselines. The decoupled approach to exploration and selection reduces vulnerability to cascading errors from suboptimal initial conditions in automated data science tasks.