Beyond Natural Language Plans: Structure-Aware Planning for Query-Focused Table Summarization
Weijia Zhang, Songgaojun Deng, Evangelos Kanoulas · 2025
Read Beyond Natural Language Plans: Structure-Aware Planning for Query-Focused Table Summarization by Weijia Zhang, Songgaojun Deng, Evangelos Kanoulas (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-22829.txt.
Introduction
Weijia Zhang, Songgaojun Deng, and Evangelos Kanoulas replace loose text plans with structured plans for query-focused table summaries, in work from twenty twenty-five. Summarising tables for a query needs multi-step reasoning over rows. Text plans are vague and hard to turn into runnable code. That flaw hurts scale, above all across many joined tables. Their structured plan, TaSoF, borrows form from multi-agent system design.
A three-phase frame, SPaGe, runs the structured reasoning end to end. Phase one writes the structured plan from the query directly. Phase two turns steps into SQL and maps their dependency graph. Parallel execution follows that graph where steps allow it. Phase three writes the final query-focused summary from results. Three public benchmarks show steady wins over prior models. Both single-table and multi-table tasks improve under the frame.
Catalog record
| CREATOR | Weijia Zhang, Songgaojun Deng, Evangelos Kanoulas |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 32 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-22829
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Beyond Natural Language Plans: Structure-Aware Planning for Query-Focused Table Summarization Weijia Zhang, Songgaojun Deng, Evangelos Kanoulas Abstract Query-focused table summarization requires complex reasoning, often approached through step-by-step natural language (NL) plans. However, NL plans are inherently ambiguous and lack structure, limiting their conversion into executable programs like SQL and hindering scalability, especially for multi-table tasks. To address this, we propose a paradigm shift to structured representations. We introduce a new structured plan, TaSoF, inspired by formalism in traditional multi-agent systems, and a framework, SPaGe, that formalizes the reasoning process in three phases: 1) Structured Planning to generate TaSoF from a query 2) Graph-based Execution to convert plan steps into SQL and model dependencies via a directed cyclic graph for parallel execution, and 3) Summary Generation to produce query-focused summaries. Our method explicitly captures complex dependencies and improves reliability. Experiments on three public benchmarks show that SPaGe consistently outperforms prior models in both single- and multi-table settings, demonstrating the advantages of structured representations for robust and scalable summarization. Figure 1: An example comparing a natural language plan with our proposed structured plan, TaSoF. The natural language plan includes vague terms such as pet ownership , which can be hard for LLMs to interpret. In contrast, TaSoF follows standard class definitions and specifies table-specific operations, making it easier for LLMs to understand. 1 Introduction Query-focused table summarization aims to generate a concise, textual summary from tabular data, specifically tailored to a user’s query. This task has recently gained significant attention, emerging as a critical challenge in the field of table understanding Zhao et al. 2023 ; Zhao et al. 2024 ; Zhang et al. 2024a . A central hurdle in this domain is table reasoning, which involves accurately extracting and aggregating query-relevant information from potentially complex input tables. Earlier studies predominantly focused on training end-to-end models that performed reasoning implicitly Liu et al. 2022b ; Liu et al. 2022a ; Jiang et al. 2022 . While these approaches showed promising results, they often required extensive training data and lacked interpretability Nguyen et al. 2025 . With the rapid advancements in large language models (LLMs), recent research has shifted towards explicit, multi-step table reasoning. These cutting-edge methods typically leverage techniques like query decomposition Ye et al. 2023 ; Zhao et al. 2024 ; Nguyen et al. 2025 or pre-defined workflows Wang et al. 2024 ; Zhang et al. 2024c ; Li et al. 2024 to generate a step-by-step reasoning plan and then execute these steps to perform the necessary table operations. However, a significant limitation persists: these reasoning plans are overwhelmingly generated as free-form natural language (NL). This introduces substantial challenges because NL plans are inherently unreliable and lack the explicit structure crucial for robust execution. This linguistic ambiguity makes the automated conversion of plan steps into executable programs for table operations, such as SQL queries, a brittle and error-prone process; even minor variations in phrasing can lead to critical execution failures Zhao et al. 2024 ; Nguyen et al. 2025 . Furthermore, the absence of a formal, machine-readable structure severely limits the scalability of these methods to complex multi-table scenarios and hinders opportunities for crucial execution optimizations. To address these limitations, we propose a paradigm shift from ambiguous NL plans to a formal, structured representation. Inspired by formalism in traditional multi-agent systems d’Inverno et al. 1997 ; Hilaire et al. 2000 ; Gruer et al. 2002 , we standardize the reasoning plan by defining it as a Ta ble S tate and o peration F low (TaSoF). In this framework, each planning step is defined as a Step class with a schema. Specifically, the Step class defines a specific table operation with its input and output states. An example of comparison between an NL plan and TaSoF is shown in Figure 1 . From this example, we observe that natural language plans often include linguistic ambiguities, such as the phrase “ pet ownership ” in the second step, which can be difficult for the LLM-based execution module to interpret accurately. In contrast, our structured plan adopts class definitions from software engineering, using explicit expressions like Has_Pet = “Yes” , which are easier to understand and translate into executable programs. Built upon this, we then present SPaGe ( S tructured P lanning a nd G raph-based e xecution), a new framework designed to robustly handle both single- and multi-table summarization scenarios. Our framework operates in three distinct phases: 1) Structured Planning : Given a complex user query, an LLM-based planner generates a hierarchical, multi-step structured plan in the TaSoF format. This key-value structure explicitly models key cross-step dependencies and is more concise and less ambiguous than its natural language counterpart. 2) Graph-based Execution : For each step within the structured plan, an executor leverages the explicit key-value structure to convert the step into a reliable executable SQL query. An external code interpreter then executes this query, yielding the intermediate output table. This structured representation enables us to model the entire plan as a Directed Acyclic Graph (DAG), facilitating the efficient parallel execution of independent steps. 3) Summary Generation : Once all plan steps have been successfully executed and intermediate results are obtained, a dedicated summary generator produces a human-readable natural language summary tailored to the user query. We thoroughly evaluate SPaGe against existing methods on three public query-focused table summarization datasets: FeTaQA, QTSumm, and QFMTS. Our experimental results consistently demonstrate that our framework significantly outperforms current baseline models, underscoring its effectiveness. Extensive analysis further reveals the superiority of our proposed structured plan format compared to natural language plans, leading to more robust and reliable execution. We also provide qualitative analysis to illustrate how our structured plan enables smoother and more accurate execution. To the best of our knowledge, our method is the first to formalize ambiguous natural language plans into a fully structured, machine-readable format specifically for query-focused table summarization. Our key contributions are summarized as follows: • We define TaSoF, a novel structure that formalizes reasoning plans as a flow of table states and operations, thereby improving execution efficiency and accuracy. • We propose SPaGe, a new structure-aware framework that leverages this structured planning and graph-based execution to achieve superior performance on the query-focused table summarization task. • Extensive experiments and analysis on three benchmarks demonstrate the effectiveness of our approach in complex multi-table reasoning and its ability to enable efficient parallel execution. 2 Related Work This section reviews two key areas of related work: program-augmented table reasoning and planning for table understanding. 2.1 Program-Augmented Table Reasoning Table reasoning necessitates models that can mimic human data analysis, encompassing operations like arithmetic calculations (e.g., counting, aggregation) and the judicious filtering of irrelevant rows or columns. This capability is a fundamental component for various table understanding tasks. Earlier approaches Liu et al. 2022b ; Jiang et al. 2022 primarily relied on end-to-end reasoning, fine-tuning black-box models to directly produce task-specific outputs. A significant limitation of these methods is their inherent lack of interpretability, making it challenging to understand the intricate intermediate reasoning steps. More recent studies Cheng et al. 2023 ; Ye et al. 2023 ; Zhang et al. 2024b have adopted program-augmented reasoning frameworks, often by employing tool-augmented LLMs Schick et al. 2023