SEA: Stateful Execution Environment for Conversational Big Data Analytics
Abstract
Applying large language model (LLM) agents to conversational data analytics is challenging, as existing agents often operate statelessly, leading to inefficiency and a fragmented user experience in multi-turn interactions. We argue that the agent's environment should explicitly encode the domain's predictable workflow. This reframes the agent's role from complex, open-ended planning to a more tractable task: strategically selecting where to resume a structured process to maximize state reuse. To this end, we introduce the Stateful Execution Environment (SEA), a framework that represents the data analysis workflow as a Directed Acyclic Graph (DAG). A key feature of SEA is its dual-representation state model, which decouples a lightweight, symbolic state graph for the LLM planner from a full computational state graph used for execution. We evaluate SEA on GloboMart, a new large-scale benchmark for conversational data analytics. Our experiments show that the planner achieves over 95\% accuracy on its reframed task, leading to an 84\% end-to-end task success rate and a 36\% reduction in average latency on stateful follow-up queries. Our work demonstrates that designing environments with strong workflow priors is a critical step toward building more efficient and reliable agents for domain-specific reasoning.