Spreadsheets are central to real-world applications such as enterprise reporting, auditing, and scientific data management. Despite their ubiquity, existing large language model based approaches typically treat tables as plain text, overlooking critical layout cues and visual semantics. Moreover, real-world spreadsheets are often massive in scale, exceeding the input length that LLMs can efficiently process. To address these challenges, we propose SpreadsheetAgent, a two-stage multi-agent framework for spreadsheet understanding that adopts a step-by-step reading and reasoning paradigm. Instead of loading the entire spreadsheet at once, SpreadsheetAgent incrementally interprets localized regions through multiple modalities, including code execution results, images, and LaTeX tables. The method first constructs a structural sketch and row/column summaries, and then performs task-driven reasoning over this intermediate representation in the Solving Stage. To further enhance reliability, we design a verification module that validates extracted structures via targeted inspections, reducing error propagation and ensuring trustworthy inputs for downstream reasoning. Extensive experiments on two spreadsheet datasets demonstrate the effectiveness of our approach. With GPT-OSS-120B, SpreadsheetAgent achieves 38.16% on Spreadsheet Bench, outperforming the ChatGPT Agent baseline (35.27%) by 2.89 absolute points. These results highlight the potential of SpreadsheetAgent to advance robust and scalable spreadsheet understanding in real-world applications. Code is available at https://github.com/renhouxing/SpreadsheetAgent.git.
Figures & tables
Figure 1: Overview of the proposed multi-agent framework for spreadsheet understanding. An Extraction Agent incrementally parses the spreadsheet with the help of code execution, vision range description, and L a T e X range description tools, while a Verification Agent cross-checks results using VLM-based and LLM-based range descriptions. Through iterative reflection and feedback, the system produces a faithful structural representation that preserves layout semantics under tight context budgets.
Model
Soft Restriction ( ↑ )
Hard Restriction ( ↑ )
Cell-Level
Sheet-Level
Overall
Cell-Level
Sheet-Level
Overall
Binder (GPT-3.5)
1.58
0.05
1.17
0.00
0.00
0.00
Mixtral-8x7B
3.39
4.67
3.88
2.32
3.71
2.85
Llama-3-70B
1.13
7.90
3.74
0.71
7.14
3.18
GPT-3.5
3.33
13.11
7.09
2.50
9.97
5.37
GPT-4o
13.49
22.51
16.96
10.52
17.66
13.27
Table 1: Main experimental results on SpreadsheetBench. We report accuracy under both Soft Restriction and Hard Restriction settings, further broken down into cell-level, sheet-level, and overall scores. Numbers for existing methods are taken from the SpreadsheetBench paper.
Model
Soft Restriction ( ↑ )
Hard Restriction ( ↑ )
Cell-Level
Sheet-Level
Overall
Cell-Level
Sheet-Level
Overall
SpreadsheetAgent
20.62
25.17
22.37
16.04
22.22
18.42
w/o Tools & Verify
19.61
21.08
20.18
14.44
18.52
16.01
w/ JSON
19.37
22.98
20.76
13.90
20.23
16.34
w/o Structure
18.84
21.08
19.70
14.08
17.66
15.46
w/o Verify
20.44
23.08
21.45
15.69
20.51
17.54
Table 2: Ablation study of SpreadsheetAgent using Qwen3-30B as the reasoning model. We examine the contributions of verification, structural sketching, and multi-format tools.
Model
Extractor
Soft
Hard
Qwen3-30B
Qwen3-480B
22.37
18.42
Qwen3-30B
Qwen3-30B
17.47
13.82
Qwen3-30B
GPT-OSS-20B
19.44
15.79
Qwen3-30B
-
16.41
12.83
GPT-OSS-20B
Qwen3-480B
32.16
29.25
GPT-OSS-20B
Qwen3-30B
25.71
20.83
Table 3: Ablation study on the impact of different structure extractors. The multi-modal reasoning model is fixed with GLM-4.5V, while we vary the language reasoning model among Qwen3-480B, Qwen3-30B, and GPT-OSS-20B.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 2: Case 1: Partial view of the leave entries spreadsheet.
Figure 3: Case 2: Partial view of the sales data spreadsheet.
Model
Fact Checking
Numerical Reasoning
Structure Comprehending
EM
F1
EM
F1
EM
F1
GPT4o
60.31
68.97
38.65
50.12
63.04
71.14
Gemini1.5-pro
59.08
66.14
35.54
43.74
63.64
69.71
DeepSeek-R1
70.91
79.45
70.31
72.54
82.71
84.62
Llama3.1-8B-Instruct
30.32
44.93
14.53
27.21
35.90
50.80
Llama3.1-8B-Instruct ∗
33.20
44.15
10.38
22.58
25.51
39.51
Appendix
Table 6: Performance comparison across different models on three sub-tasks of RealHiT Bench: Fact Checking, Numerical Reasoning, and Structure Comprehending. Results marked with ∗ indicate re-evaluations under a slightly modified prompt that encourages models to perform explicit reasoning before producing final answers, as opposed to directly outputting results.
Model
Fact Checking
Numerical Reasoning
Structure Comprehending
EM
F1
EM
F1
EM
F1
GPT-OSS-20B
42.81
48.54
34.63
39.03
33.84
40.46
w/ TreeThinker
45.11
50.12
39.82
44.42
33.59
39.53
w/ SpreadsheetAgent
48.73
53.64
44.62
48.92
39.65
44.74
GPT-OSS-120B
58.18
62.75
55.38
59.16
54.04
57.19
w/ TreeThinker
56.53
61.14
54.60
59.06
53.03
57.98
Appendix
Table 7: Performance comparison across different models with Python tool on three sub-tasks of RealHiT Bench: Fact Checking, Numerical Reasoning, and Structure Comprehending.