XL-2: A Foundation Model for Spreadsheet Intelligence

Abstract

We introduce XL-2, a foundation model designed specifically for structured financial spreadsheet generation and comprehension. Unlike general-purpose language models that treat spreadsheets as flat text, XL-2 encodes the inherent graph structure of cell references, formula dependencies, and cross-sheet links. Trained on 1.2M anonymized institutional real estate models, XL-2 achieves leading results on financial model generation, assumption extraction, and formula prediction tasks — outperforming GPT-4 by 34% on our CRE-Bench benchmark and producing audit-ready .xlsx output with 97.3% formula accuracy.

The Problem with LLMs and Spreadsheets

Large language models have demonstrated remarkable capability across text, code, and reasoning tasks. However, financial spreadsheets present a fundamentally different challenge. A typical institutional real estate acquisition model contains 2,000-8,000 cells across 6-12 sheets, with complex formula dependency graphs, conditional logic, circular references (for debt sizing), and implicit structural conventions that vary by asset class, deal type, and firm preference.

When prompted to generate a spreadsheet, current LLMs produce output that looks correct — the labels are right, the numbers are plausible — but fails under audit. The most common failure modes are:

Table 1 — LLM failure modes in financial model generation, measured across 500 GPT-4 generated acquisition models evaluated by institutional analysts.

These aren't edge cases. In our evaluation, 83% of GPT-4 generated models contained at least one error that would be caught in an institutional IC review. The fundamental issue is representational: LLMs serialize spreadsheets as text, losing the spatial and relational structure that makes a spreadsheet a spreadsheet.

Architecture

XL-2 addresses this with a tri-modal architecture that jointly encodes three distinct signal types:

Cell Graph Encoder. Each cell is represented as a node in a directed graph, where edges encode formula references (=B12*C4 creates edges from B12 and C4 to the formula cell). Sheet boundaries, named ranges, and cross-sheet links are encoded as edge attributes. This preserves the structural relationships that flat text representations destroy.

Document Context Encoder. Deal documents (OMs, rent rolls, appraisals) are processed through a standard transformer encoder. The key innovation is the assumption linking layer — a cross-attention mechanism that learns to map extracted document values to specific cell positions in the target model. When the OM states "Year 1 NOI: $2.4M," the model learns not just to extract the number but to place it in the correct cell with the correct formula dependencies.

Numeric Tokenizer. Financial values span 10+ orders of magnitude ($0.50/SF to $2.4B AUM) and require decimal precision. Standard text tokenization handles this poorly. We use a specialized tokenizer that decomposes floats into (sign, exponent, mantissa) triples, allowing the model to reason about numeric magnitude independently from precision.

Table 2 — XL-2 component breakdown. Total parameter count: 1.385 billion.

Training Data

XL-2 was trained on 1.2 million anonymized financial models sourced from institutional real estate transactions. The dataset spans:

Table 3 — Training data distribution by asset class. LIHTC models are disproportionately complex due to tax credit layering and compliance requirements.

All models were anonymized at the entity level (property names, addresses, investor identities replaced with synthetic equivalents) while preserving structural and numeric relationships. The anonymization pipeline is described in Appendix B of the full paper.

Benchmarks

We evaluate on CRE-Bench, a new benchmark we're releasing alongside this paper. CRE-Bench contains 2,400 model generation tasks across 8 categories, each with human-validated ground truth and automated evaluation metrics.

Table 4 — CRE-Bench results by task category. XL-2 approaches human expert performance across all categories and significantly outperforms general-purpose LLMs.

The gap between XL-2 and general-purpose LLMs is widest on waterfall distributions and LIHTC underwriting — precisely the structures that require deep domain knowledge and complex conditional logic.

Case Study: 240-Unit Multifamily Acquisition

To illustrate XL-2 in practice, we walk through a complete acquisition underwriting for a 240-unit Class B multifamily property in Austin, TX. The input is a 47-page offering memorandum.

Key outputs from the generated model:

Table 5 — XL-2 vs. human analyst output comparison. Variances stem from rounding differences in expense growth assumptions (XL-2 used monthly compounding; the analyst used annual).

The 0.18% variance in Year 1 NOI traces to a single assumption: XL-2 compounded the 3% expense growth rate monthly, while the analyst applied it annually. Both approaches are defensible. The remaining metrics cascade from this difference.

Limitations and Open Questions

XL-2 has clear limitations that we want to be transparent about:

Novel structures. XL-2 performs best on deal structures well-represented in its training data. For genuinely novel structures — a first-of-its-kind public-private partnership, an unusual tax credit stacking arrangement — performance degrades. The model produces output, but it requires more analyst review.

Judgment calls. XL-2 can extract assumptions from documents and apply them consistently, but it cannot make the judgment calls that distinguish good underwriting from great underwriting. Choosing a reversion cap rate, sizing a renovation budget, or deciding whether a sponsor's track record justifies a promote structure — these remain human decisions.

Data freshness. The training data reflects market conditions through Q4 2025. Cap rate assumptions, rent growth expectations, and construction cost benchmarks embedded in the model's priors will drift as markets evolve. We are implementing continuous fine-tuning on new transaction data.

Conclusion

XL-2 demonstrates that purpose-built models for structured financial documents significantly outperform general-purpose LLMs on real-world institutional tasks. The key insight is representational: by encoding spreadsheet structure as a graph rather than serialized text, XL-2 preserves the relationships that make financial models useful — and auditable.

The full paper, CRE-Bench dataset, and evaluation code are available at apers.app/research/xl2. We welcome feedback from the institutional real estate community.