跳转至

文章背景与核心概要

大语言模型（LLM）在处理大规模和复杂的表格时，常常因为无关上下文的干扰以及难以定位必要证据而表现不佳。为了解决这一局限性，本文提出了 PARTAB（Partition-Aware Reasoning over Tables，表格分区感知推理）框架。PARTAB 作为 LLM 与表格之间的结构化证据接口，将与查询相关的证据表示为语义连贯、行链接的表格区域。

该技术通过对列组和行级分区进行分层选择过程，组合出有针对性的证据，从而显著提升了推理准确性，减少了上下文开销，并能够高效扩展到诸如问答和事实核查等复杂的基于表格的任务中。本文在多个表格推理基准上对 PARTAB 进行了评估，证明了结构化、分区感知的证据构建对于可扩展表格推理的重要价值。

PARTAB：基于结构化证据和分区感知推理的可扩展表格理解 (PARTAB: Partition-Aware Reasoning with Structured Evidence for Scalable Table Understanding)

摘要 (Summary)

Large Language Models (LLMs) often struggle with large and complex tables due to irrelevant context and the difficulty of localizing necessary evidence. To address this limitation, the paper introduces PARTAB (Partition-Aware Reasoning over Tables). PARTAB acts as a structured evidence interface between the LLM and the table by representing query-relevant evidence as semantically coherent, row-linked table regions. Using a hierarchical selection process over column groups and row-level partitions, PARTAB composes targeted evidence to significantly improve reasoning accuracy, reduce context overhead, and scale efficiently across complex table-based tasks like question answering and fact verification.

Large Language Models (LLMs) often struggle with large and complex tables due to irrelevant context and the difficulty of localizing necessary evidence. To address this limitation, the paper introduces PARTAB (Partition-Aware Reasoning over Tables). PARTAB acts as a structured evidence interface between the LLM and the table by representing query-relevant evidence as semantically coherent, row-linked table regions. Using a hierarchical selection process over column groups and row-level partitions, PARTAB composes targeted evidence to significantly improve reasoning accuracy, reduce context overhead, and scale efficiently across complex table-based tasks like question answering and fact verification.

论文元数据 (Paper Metadata)

  • arXiv ID: arXiv:2608.24082 [cs.CL]
  • 学科分类 (Subjects): 计算与语言 (cs.CL)；人工智能 (cs.AI)；信息检索 (cs.IR)
  • 作者 (Authors): Md Mahadi Hasan Nahid, Davood Rafiei
  • 提交时间 (Submitted On): 2026年8月25日
  • 篇幅与图表 (Length): 21 页，13 张图

摘要正文 (Abstract)

大语言模型（LLM）在表格推理方面表现出了强大的能力，但随着表格规模和复杂性的增长，其有效性会因无关上下文以及难以定位推理所需证据而下降。现有的方法通常对整个表格或单一简化视图进行推理，这仍然可能掩盖重要的行-列关系。我们引入了 PARTAB（Partition-Aware Reasoning over Tables），这是一个在 LLM 和表格之间构建结构化证据接口的框架。PARTAB 将与查询相关的证据表示为语义连贯、行链接的表格区域，并在列组和行级分区上执行分层选择，然后组合所选证据以生成答案。我们在多个表格推理基准（涵盖问答、事实核查和数值推理）上评估了 PARTAB。PARTAB 持续超越了全表提示词和几种最近的表格推理方法，在 WikiTableQuestions 和 TabFact 上取得了强大的性能，同时在数值推理上保持了竞争力。进一步的分析表明，语义分区和有针对性的证据选择改善了证据定位，大幅减少了推理上下文，并对复杂表格产生了更大的效益。这些结果证明了结构化、分区感知的证据构建对于可扩展表格推理的价值。

Large Language Models (LLMs) have shown strong capabilities in table reasoning, but their effectiveness degrades as tables grow in size and complexity due to irrelevant context and difficulty localizing the evidence required for reasoning. Existing approaches typically reason over either the full table or a single reduced view, which can still obscure important row-column relationships. We introduce PARTAB (Partition-Aware Reasoning over Tables), a framework that constructs a structured evidence interface between the LLM and the table. PARTAB represents query-relevant evidence as semantically coherent, row-linked table regions and performs hierarchical selection over column groups and row-level partitions before composing the selected evidence for answer generation. We evaluate PARTAB on multiple table reasoning benchmarks, covering question answering, fact verification, and numerical reasoning. PARTAB consistently improves over full-table prompting and several recent table reasoning methods, achieving strong performance on WikiTableQuestions and TabFact while remaining competitive on numerical reasoning. Additional analyses show that semantic partitioning and targeted evidence selection improve evidence localization, substantially reduce the reasoning context, and provide larger benefits on complex tables. These results demonstrate the value of structured, partition aware evidence construction for scalable table reasoning.

访问与资源 (Access & Resources)