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

Data science agents often rely on a single initial plan, making them highly susceptible to cascading errors if that initial state is suboptimal. The authors investigate whether test-time scaling—generating multiple initial states and selecting a subset for parallel execution—can mitigate these errors. Specifically, they ask how decoupling the generation of these states from their selection influences agent performance across both closed-ended and open-ended data science tasks.

The authors introduce CIPHER, a framework that separates the process into three distinct phases: (1) generating a broad set of candidate initial states, (2) selecting a subset of these candidates using strategies like clustering or entropy maximization, and (3) executing these states in parallel before aggregating the results. Unlike previous approaches that conflate generation and selection, CIPHER allows for independent control over the number of candidates generated (N) and the number of candidates executed (M). The authors evaluate this framework on two benchmarks: Infi-DA-Bench (closed-ended) and InsightBench (open-ended).

CIPHER consistently outperforms state-of-the-art agents in matched-model comparisons. The study reveals that increasing the selection budget (M) provides reliable performance gains across all task types, whereas increasing the generation budget (N) yields diminishing returns that vary by task complexity. A critical finding is that the benefits of the decoupled framework are highly dependent on the aggregator model; using a more powerful 'leader' model to synthesize the parallel outputs significantly improves performance compared to using the same base model for aggregation. This suggests that weaker models may lack the capacity to effectively exploit the diversity provided by the exploration-selection process.