PExA: Parallel Exploration Agent for Complex Text-to-SQL

Researchers led by Tanmay Parekh have introduced PExA (Parallel Exploration Agent), a new framework for complex text-to-SQL tasks that reimagines query generation through the lens of software test coverage. Accepted at ACL 2026, PExA tackles the persistent latency-performance trade-off in LLM-based agents by preparing the original query with a suite of test cases composed of simpler, atomic SQLs. These atomic SQLs are executed in parallel, together ensuring semantic coverage of the original query. The final SQL is generated only after iterating on test case coverage and gathering enough information, grounding the output in explored test case SQLs.

Validated on the state-of-the-art Spider 2.0 benchmark, PExA achieved a new state-of-the-art with 70.2% execution accuracy. This parallel exploration approach allows the agent to maintain high performance without the latency penalties typical of sequential methods. By breaking down complex queries into parallel testable units, PExA offers a scalable path for enterprise applications requiring real-time natural language to database query translation.