Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SheetBrain: A Neuro-Symbolic Agent for Accurate Reasoning over Complex and Large Spreadsheets

Ziwei Wang, Jiayuan Su, Mengyu Zhou, Huaxing Zeng, Mengni Jia, Xiao Lv, Haoyu Dong, Xiaojun Ma, Shi Han, Dongmei Zhang

TL;DR

This work tackles the challenge of reasoning over complex spreadsheets with large language models by introducing SheetBrain, a neuro-symbolic agent built around an understand-execute-validate pipeline. The Understanding Module provides a global, query-aware overview with preserved spatial context; the Execution Module uses a Python sandbox, Excel tooling, and a symbolic dataflow architecture to perform precise, multi-turn reasoning; and the Validation Module enforces quality assurance with iterative re-execution. SheetBench is proposed as a challenging benchmark for large, multi-table, and structurally complex spreadsheets, and SheetBrain achieves state-of-the-art performance on public QA and manipulation datasets as well as SheetBench, aided by ablations that demonstrate the importance of understanding, validation, and a robust tool-suite. The results underscore the value of symbolic dataflow and structured, global verification for reliable spreadsheet reasoning and manipulation, enabling more robust deployments in real-world enterprise contexts.

Abstract

Understanding and reasoning over complex spreadsheets remain fundamental challenges for large language models (LLMs), which often struggle with accurately capturing the complex structure of tables and ensuring reasoning correctness. In this work, we propose SheetBrain, a neuro-symbolic dual workflow agent framework designed for accurate reasoning over tabular data, supporting both spreadsheet question answering and manipulation tasks. SheetBrain comprises three core modules: an understanding module, which produces a comprehensive overview of the spreadsheet - including sheet summary and query-based problem insight to guide reasoning; an execution module, which integrates a Python sandbox with preloaded table-processing libraries and an Excel helper toolkit for effective multi-turn reasoning; and a validation module, which verifies the correctness of reasoning and answers, triggering re-execution when necessary. We evaluate SheetBrain on multiple public tabular QA and manipulation benchmarks, and introduce SheetBench, a new benchmark targeting large, multi-table, and structurally complex spreadsheets. Experimental results show that SheetBrain significantly improves accuracy on both existing benchmarks and the more challenging scenarios presented in SheetBench. Our code is publicly available at https://github.com/microsoft/SheetBrain.

SheetBrain: A Neuro-Symbolic Agent for Accurate Reasoning over Complex and Large Spreadsheets

TL;DR

This work tackles the challenge of reasoning over complex spreadsheets with large language models by introducing SheetBrain, a neuro-symbolic agent built around an understand-execute-validate pipeline. The Understanding Module provides a global, query-aware overview with preserved spatial context; the Execution Module uses a Python sandbox, Excel tooling, and a symbolic dataflow architecture to perform precise, multi-turn reasoning; and the Validation Module enforces quality assurance with iterative re-execution. SheetBench is proposed as a challenging benchmark for large, multi-table, and structurally complex spreadsheets, and SheetBrain achieves state-of-the-art performance on public QA and manipulation datasets as well as SheetBench, aided by ablations that demonstrate the importance of understanding, validation, and a robust tool-suite. The results underscore the value of symbolic dataflow and structured, global verification for reliable spreadsheet reasoning and manipulation, enabling more robust deployments in real-world enterprise contexts.

Abstract

Understanding and reasoning over complex spreadsheets remain fundamental challenges for large language models (LLMs), which often struggle with accurately capturing the complex structure of tables and ensuring reasoning correctness. In this work, we propose SheetBrain, a neuro-symbolic dual workflow agent framework designed for accurate reasoning over tabular data, supporting both spreadsheet question answering and manipulation tasks. SheetBrain comprises three core modules: an understanding module, which produces a comprehensive overview of the spreadsheet - including sheet summary and query-based problem insight to guide reasoning; an execution module, which integrates a Python sandbox with preloaded table-processing libraries and an Excel helper toolkit for effective multi-turn reasoning; and a validation module, which verifies the correctness of reasoning and answers, triggering re-execution when necessary. We evaluate SheetBrain on multiple public tabular QA and manipulation benchmarks, and introduce SheetBench, a new benchmark targeting large, multi-table, and structurally complex spreadsheets. Experimental results show that SheetBrain significantly improves accuracy on both existing benchmarks and the more challenging scenarios presented in SheetBench. Our code is publicly available at https://github.com/microsoft/SheetBrain.
Paper Structure (43 sections, 12 figures, 6 tables)

This paper contains 43 sections, 12 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Understanding Module
  3. Execution Module
  4. Excel-Specific Tooling Protocol
  5. Symbolic Dataflow Architecture
  6. Iterative Reasoning and Execution Cycle
  7. Validation Module
  8. SheetBench
  9. Experiments and Analysis
  10. Experimental Setup
  11. Benchmarks
  12. Baselines
  13. Evaluator
  14. Main Results
  15. QA
  16. ...and 28 more sections

Figures (12)

  • Figure 1: SheetBrain Pipeline. SheetBrain adopts a robust three-stage pipeline: understand, execute, and validate. It first constructs a global, query-aware understanding of the spreadsheet, then performs tool-augmented reasoning in a Python sandbox, and finally validates and refines its output through iterative self-correction.
  • Figure 2: The workflow of the understanding module.
  • Figure 3: Statistics of SheetBench. The numbers shown in the pie chart represent the counts for each data type.
  • Figure 4: Case 1 Sheet
  • Figure 6: Case 2 Sheet
  • ...and 7 more figures