LLM‑powered data analytics aims to turn natural‑language instructions into executable analytical programs over underlying data. Existing approaches either treat LLMs as intent‑to‑code translators or as standalone analytical engines, but they often struggle to align domain semantics with complex schemas and heterogeneous metadata, leading to brittle execution and inconsistent results.
LYNX formulates LLM‑driven analytics as a schema‑ and knowledge‑constrained analysis problem instead of a single‑pass generation task. The system organizes synthesis into a closed‑loop workflow—planning, generation, verification, and refinement—through three role‑specialized agents that enforce schema grounding, semantic consistency, and execution correctness.
To mitigate domain ambiguity and schema misalignment, LYNX integrates Semantic‑Structural Distillation and Stage‑Aware Allocation (SSD‑SAA). A delta‑guided iterative revision continuously diagnoses and repairs semantic and execution‑level failures, markedly improving robustness.
Evaluations on the NL‑to‑SQL benchmarks BIRD‑SQL Mini‑Dev and Spider 2.0‑Lite show execution accuracy ranging from 10.66% to 59.54% across lightweight and commercial models, while reducing token consumption by 70.53% compared with the prior most efficient method. The framework generalizes well from open‑source lightweight backbones to the latest commercial models.
Case studies on quantitative finance analytics and complex multi‑relational SQL workloads further demonstrate the system’s robustness, scalability, and practical applicability in real‑world scenarios.
Review