Recent advances in multimodal Retrieval-Augmented Generation (RAG) enable Large Language Models (LLMs) to analyze enterprise spreadsheet workbooks containing millions of cells, cross-sheet dependencies, and embedded visual artifacts. However, state-of-the-art approaches exclude critical context through single-pass retrieval, lose data resolution through compression, and exceed LLM context windows through naive full-context injection, preventing reliable multi-step reasoning over complex enterprise workbooks. We introduce Beyond Rows to Reasoning (BRTR), a multimodal agentic framework for spreadsheet understanding that replaces single-pass retrieval with an iterative tool-calling loop, supporting end-to-end Excel workflows from complex analysis to structured editing. Supported by over 200 hours of expert human evaluation, BRTR achieves state-of-the-art performance across three frontier spreadsheet understanding benchmarks, surpassing prior methods by 25 percentage points on FRTR-Bench, 7 points on SpreadsheetLLM, and 32 points on FINCH. We evaluate five multimodal embedding models, identifying NVIDIA NeMo Retriever 1B as the top performer for mixed tabular and visual data, and vary nine LLMs. Ablation experiments confirm that the planner, retrieval, and iterative reasoning each contribute substantially, and cost analysis shows GPT-5.2 achieves the best efficiency-accuracy trade-off. Throughout all evaluations, BRTR maintains full auditability through explicit tool-call traces.
Figures & tables
Figure 1: Overview of the BRTR framework pipeline: multimodal spreadsheet indexing, agentic task planning and decomposition, specialized tool execution, and multi-format response generation with full end-to-end tool-trace auditability.
Figure 2: Conceptual illustration of answer accuracy as a function of spreadsheet size. BRTR maintains near-perfect accuracy across all scales by iteratively re-querying until evidence is sufficient, while single-pass and compression-based methods degrade as workbook complexity increases. Naïve full-context approaches exceed LLM context windows beyond 50K cells. Trends are derived from aggregate observations across evaluations; see Tables 2 and 3 for precise measurements.
K=5
K=10
Embedding Model
nDCG
Recall
MAP
nDCG
Recall
MAP
Amazon Titan
0.26
0.39
0.27
0.33
0.47
0.27
Cohere v4
0.33
0.45
0.35
0.40
0.51
0.34
Amazon Nova
0.30
0.45
0.32
0.40
0.58
0.33
NVIDIA NeMo
0.33
0.47
0.33
0.42
0.60
0.34
Jina v4
0.19
0.26
0.19
0.23
0.33
0.19
Table 1: Multimodal embedding model comparison on FRTR-Bench (146 queries, 28 files). Bold indicates best per metric.
Accuracy
BRTR Operational Cost
Model
BRTR
FRTR
SpreadsheetLLM
Claude in Excel
Tool Calls
Tokens
Latency (s)
GPT-5
0.99
0.73
0.18
–
4.01
68,061
166.34
Gemini 3 Pro
0.99
0.68
0.29
–
5.40
48,462
217.77
Claude Opus 4.6
0.98
0.74
0.34
0.84
5.20
462,614
143.23
GPT-5.2
0.98
0.72
0.32
–
3.05
20,127
90.81
Claude Sonnet 4.5
0.96
0.74
0.13
0.69
6.20
92,202
163.83
Table 2: Cross-technique performance on FRTR-Bench (159 queries, 30 files). BRTR achieves up to 99% accuracy, 25 percentage points above single-pass FRTR. FRTR and SpreadsheetLLM use mean context sizes of 7,691 and 12,745 tokens respectively. Claude in Excel supports only Claude models. Operational cost columns report means. Bold indicates highest accuracy per technique.
Figure 3: Answer accuracy on FRTR-Bench across nine LLMs and four techniques. BRTR (teal) consistently dominates, achieving up to 99% accuracy with frontier models, 25 percentage points above the best single-pass FRTR baseline. SpreadsheetLLM compression (lavender) performs poorly throughout, confirming that cross-sheet references require retrieval rather than compression alone.
Accuracy
BRTR Cost
Model
BRTR
FRTR
SpreadsheetLLM
Naïve Full-Context
Tokens
Latency (s)
GPT-5
0.98
0.87
0.90
0.68
156,306
273.36
Claude Sonnet 4.5
0.97
0.84
0.91
0.55
166,117
280.59
Gemini 2.5 Pro
0.79
0.85
0.89
0.67
27,675
223.79
GPT-4o
0.56
0.77
0.78
0.35
256,523
315.93
LLaMA 4 Maverick 17B
0.30
0.53
0.73
0.41
18,608
408.73
Table 3: Cross-technique performance on the SpreadsheetLLM benchmark (298 queries). BRTR achieves 97–98% accuracy with frontier models, while SpreadsheetLLM compression reaches 89–91% and naïve full-context approaches yield only 55–68%. FRTR, SpreadsheetLLM, and Naïve Full-Context use mean context sizes of 6,920, 5,811, and 13,631 tokens respectively. Operational cost columns report means. Bold indicates highest accuracy per technique. Models sorted by BRTR accuracy.
Accuracy
BRTR Cost (Median)
Model
BRTR
Naïve
Calls
Tok. (K)
Lat. (min)
Claude Opus 4.6
0.95
0.63
37
358
4.8
GPT-5.2
0.95
0.46
33
223
4.3
Claude Sonnet 4.5
0.87
0.20
25
305
2.7
Gemini 3 Pro
0.73
0.27
28
380
7.9
LLaMA 4 Maverick 17B
0.12
0.01
8
113
0.5
Table 4: Cross-technique performance on the FINCH benchmark (172 workflows). Naïve denotes single-call API-based agents; results for Claude Sonnet 4.5 and Gemini 3 Pro are reported by Dong et al. (2025) . Operational cost columns report medians to account for heavy-tailed workflow distributions. Bold indicates highest accuracy per technique. Models sorted by BRTR accuracy.
Figure 4: Ablation study on a 50-task FINCH subset (Claude Opus 4.6). Each bubble represents one configuration; position encodes accuracy (x) and latency (y), while bubble area encodes median token consumption. BRTR (Full) achieves the highest accuracy with 40% fewer tokens than the monolithic agent (No Planner), demonstrating that task decomposition improves both accuracy and efficiency.
Configuration
Accuracy
Tool Calls
Tokens (K)
Latency (s)
BRTR (Full)
0.88
33
312
248
No Planner
0.68
25
521
175
Naïve
0.46
–
438
190
Full Context
0.52
16
583
114
No Agent Loop
0.00
1
11
5
Table 5: Ablation study on a representative 50-task FINCH subset (Claude Opus 4.6). Each row removes one or more BRTR components. Tool Calls, Tokens, and Latency report medians. Bold indicates highest accuracy.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Model (Accuracy)
Executor
Invocations
Tasks
Tool Calls
Claude Opus 4.6 (95.3%)
Search
264
148
1,623
Excel
232
166
5,351
IO
16
14
271
Web
18
11
651
GPT-5.2 (94.8%)
Search
197
145
1,164
Excel
425
164
5,330
Appendix
Table 6: FINCH executor utilization per model. Invocations counts sub-tasks dispatched to each executor; Tasks counts distinct workflows using that executor; Tool Calls counts individual tool invocations within the executor. LLaMA 4 Maverick 17B omitted due to insufficient valid planner outputs.
Category ( n )
BRTR
− Planner
− Retrieval
− Iteration
Calculation (35)
0.94
0.69
0.54
0.00
Structuring / Formatting (23)
0.87
0.74
0.48
0.00
Cross-sheet Retrieval (14)
0.86
0.43
0.43
0.00
Validation / Review (11)
0.82
0.64
0.45
0.00
Data Entry / Import (11)
0.91
0.64
0.45
0.00
Summary / Visualization (10)
0.90
0.70
0.50
0.00
Appendix
Table 7: Ablation accuracy by task category on the 50-task FINCH subset (Claude Opus 4.6). Each task may span multiple categories. n denotes the number of tasks involving each category.
Figure 5: Example BRTR tool-call trace for a representative query. The agent iteratively invokes search tools, inspects returned chunks, refines queries, and cross-references sheets before synthesizing a grounded answer with full provenance.