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

2026-07-15

AI Summary

Current data science agents, constrained by reliance on a single initial state, are highly susceptible to early errors and struggle to perform effectively across both closed-ended and open-ended tasks. This work proposes CIPHER, a novel framework that introduces, for the first time, a decoupled exploration–selection (DES) mechanism. At test time, CIPHER generates multiple initial states in parallel and explicitly separates the generation of candidate states from their selection strategy. Integrating multi-initial-state generation, strategic selection, parallel execution architecture, and a lightweight aggregation model, this design substantially enhances agent robustness and generalization. Experiments demonstrate that CIPHER outperforms state-of-the-art methods of comparable scale on both task types and achieves performance on par with significantly larger models using a more compact base architecture.

Abstract

Data science tasks span from closed-ended information extraction to open-ended analysis, presenting significant challenges for automation. Recent AI agents powered by language models show promise for handling such complex tasks. However, existing agents typically rely on a single initial state that conditions the entire agent's execution, making them vulnerable to cascading errors initiated by a suboptimal initial state. To mitigate this, we present CIPHER, an automated data science agent that leverages test-time scaling through the generation and selection of multiple initial states for concurrent execution. Unlike existing works on test-time scaling of AI agents, CIPHER explicitly decouples the generation of candidate initial states from their strategic selection for parallel execution. Through extensive evaluation on two benchmarks (closed-form and open-form tasks), we demonstrate that CIPHER exceeds state-of-the-art performance in matched-model comparisons, and remains competitive against larger-model baselines despite relying on a substantially smaller base LM. Our empirical study characterizes the design space of the Decoupled Exploration-Selection (DES) framework: we quantify how generation strategy, selection strategy, and aggregator model capacity contribute to overall performance, and derive actionable design recommendations for practitioners.

Problem

• data science agents
• test-time scaling
• initial state selection
• cascading errors
• automation

Innovation

• Decoupled Exploration-Selection
• Test-Time Scaling
• Data Science Agents
• Initial State Generation
• Parallel Execution