FACTS: Fast, Accurate, and Privacy-Compliant Table Summarization via Offline Template Generation

Ye Yuan, Mohammad Amin Shabani, Siqi Liu

Keywords: Query-focused Table Summarization, LLMs, Agentic Frameworks

Abstract: Query-focused table summarization aims to generate concise textual summaries of tabular data conditioned on a specific user query, enabling users to access relevant insights over large tables. However, existing approaches struggle to meet real-world requirements: fine-tuned LLMs, retrieval-augmented generation (RAG) pipelines, and direct LLM summarization face token-limit constraints, high computational costs, and privacy risks, while decomposition-based strategies and manual template design are labor-intensive and non-scalable. To address these challenges, we introduce FACTS, a Fast, Accurate, and Privacy-Compliant Table Summarization approach via Offline Template Generation. FACTS leverages LLM-based agentic workflows to automatically generate offline templates, where each offline template consists of schema-aware SQL queries and a Jinja$2$ template derived from a given user query and table schema. Once generated, these offline templates can be efficiently reused across new tables, making summarization both scalable and cost-effective. Our framework achieves three advantages: fast summarization through lightweight SQL execution and reusable offline templates which avoid repeated LLM inference, accurate summaries by grounding outputs in precise query results rather than free-form text generation, and privacy-compliant deployment since only table schema is exposed to the LLM. We benchmark FACTS on QTSumm and QFMTS, demonstrating promising improvements over selected baselines.

Submission Number: 23