AI Unearths Colonial History: LLMs Outperform Humans in Georeferencing Virginia Land Grants

New research reveals how large language models transform historical archives into actionable geographic data, achieving unprecedented accuracy in locating colonial settlements.

For decades, historians have struggled with Virginia's 17th and 18th-century land patents—rich narrative descriptions of property boundaries that resisted automated spatial analysis. These "metes and bounds" documents, while poetically detailed, remained locked in textual formats. Now, a groundbreaking study demonstrates how large language models (LLMs) can decode these historical records with startling precision, outperforming both human experts and specialized geocoding tools.

The Colonial Mapping Challenge

Virginia's colonial land grants present a unique computational problem. Unlike modern geospatial data, these documents describe boundaries through natural landmarks ("a white oak tree near the creek"), adjacent landowners ("bordering Mr. Jefferson's tract"), and directional measurements ("200 poles northwest"). Traditional geoparsing tools like Stanford NER and Mordecai-3 struggle with such archaic language and contextual dependencies. Manual conversion by GIS analysts is painstakingly slow—typically costing hundreds of dollars per grant with variable accuracy.

Ryan Mioduski's research addresses this gap by introducing a rigorously verified benchmark: 5,471 digitized patent abstracts (1695-1732) with 43 precisely geolocated test cases. This dataset becomes the proving ground for evaluating LLMs' ability to bridge historical narratives and modern coordinates.

Methodology: LLMs vs. Traditional Approaches

The study tested six OpenAI models across three architectures under two distinct paradigms:

1. Direct Coordinate Extraction: Models output latitude/longitude directly from patent text
2. Tool-Augmented Chain-of-Thought: Models invoke external geocoding APIs during reasoning

These approaches faced off against four baselines: human GIS analysts, Stanford NER, Mordecai-3, and a simple county-centroid heuristic. Performance was measured by mean and median distance error in kilometers.

Results: AI's Mapping Breakthrough

The findings challenge conventional wisdom about historical georeferencing:

• Top-performing LLM (o3-2025-04-16) achieved 23 km mean error (median 14 km)—37.5% better than median LLM performance and 67% more accurate than human analysts
• Ensemble approach (5 LLM calls) reduced errors to 19 km mean (median 12 km) at just $0.20 per grant
• Cost-efficient model (gpt-4o-2024-08-06) maintained 28 km accuracy at $1.09 per 1,000 grants—orders of magnitude cheaper than manual processing
• Surprise finding: External geocoding APIs provided no measurable benefit, suggesting LLMs internalize sufficient geographic knowledge

"The patentee-name-redaction experiment proved crucial," notes the study. "When we removed personal names, error increased just 9%—confirming models rely on landscape descriptions rather than memorized records."

Implications for Historical Research

This breakthrough extends far beyond colonial Virginia:

1. Scalability: Processing thousands of documents becomes economically feasible, enabling continent-scale settlement pattern analysis
2. New methodologies: Historians can quantitatively study land distribution, indigenous displacement, and environmental changes
3. Archival democratization: Small institutions gain access to sophisticated georeferencing previously requiring specialized expertise

The New Geospatial Benchmark

The research establishes gpt-4o-2024-08-06 as a cost/accuracy benchmark for historical geoparsing. More significantly, it demonstrates how LLMs can internalize complex spatial relationships from text—a capability with applications ranging from urban archaeology to legal boundary disputes.

As the paper concludes: "We're not just mapping land grants—we're teaching AI to read historical landscapes." This fusion of natural language understanding and geospatial reasoning may soon unlock centuries of archival knowledge previously hidden in plain text.

Source: Benchmarking Large Language Models for Geolocating Colonial Virginia Land Grants by Ryan Mioduski (arXiv:2508.08266).