💬 NLP

ReaGeo: Reasoning-Enhanced End-to-End Geocoding with LLMs

This paper introduces ReaGeo, an end-to-end geocoding framework that leverages large language models to transform coordinate prediction into a text generation task enhanced by Chain-of-Thought reasoning and reinforcement learning, thereby overcoming the limitations of traditional multi-stage retrieval methods to accurately handle diverse address and spatial queries.

Original authors: Jian Cui, Zhiyuan Ren, Desheng Weng, Yongqi Zhao, Gong Wenbin, Yu Lei, Zhenning Dong

Published 2026-04-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jian Cui, Zhiyuan Ren, Desheng Weng, Yongqi Zhao, Gong Wenbin, Yu Lei, Zhenning Dong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to give directions to a friend who has never been to your city. You say, "Meet me at the big red house two blocks north of the old library."

Traditional geocoding (the old way computers do this) is like a very strict librarian. It has a massive, perfectly organized card catalog of every address. When you give it your sentence, it tries to:

Break it down: Find the "library" and the "red house" in its catalog.
Match it: Look up those exact words in the database.
Calculate: Do some math to figure out where "two blocks north" is.

The problem? If you say "the big red house" but the catalog only says "123 Main St," or if you misspell "library" as "libary," the librarian gets confused, gives up, or points you to the wrong place. It's a fragile, multi-step process where one mistake ruins the whole trip.

ReaGeo (the new method in this paper) is like a local expert who grew up in that city. They don't need a card catalog. They just know the neighborhood.

Here is how ReaGeo works, broken down into simple concepts:

1. The Secret Code: Geohash

Instead of trying to guess exact GPS coordinates (like 40.7128° N, 74.0060° W), ReaGeo translates the map into a secret code called a "Geohash."

The Analogy: Imagine the world is a giant chocolate bar. Geohash is like snapping off smaller and smaller squares of chocolate.
- Level 1: A huge continent-sized square.
- Level 5: A city-sized square.
- Level 9: A tiny square about the size of a living room (2.4 meters).
ReaGeo doesn't "calculate" the math; it writes the code for that tiny living-room-sized square, just like a human would write a word.

2. The "Think Aloud" Strategy (Chain-of-Thought)

Sometimes the address is vague, like "near the park." A standard computer might guess randomly. ReaGeo uses a technique called Chain-of-Thought (CoT).

The Analogy: Before ReaGeo gives you the answer, it whispers its thinking process out loud.
- Input: "Meet me 200 meters south of the park."
- ReaGeo's Whisper: "Okay, I know where the park is. South means down on the map. 200 meters is about two football fields. So, I need to move down two football fields from the park's edge."
- Output: The secret Geohash code for that spot.
This helps the model understand the relationship between things, not just the words themselves.

3. The Coach with a Whistle (Reinforcement Learning)

Training a model is like teaching a dog to fetch. If the dog brings the stick, you give it a treat. If it brings a rock, you say "no."

ReaGeo is trained using Reinforcement Learning.
The Analogy: Every time ReaGeo guesses a location, a "Coach" measures the distance between the guess and the real location.
- If the guess is 500 meters away, the Coach gives a "bad score."
- If the guess is 5 meters away, the Coach gives a "great score."
Over millions of tries, ReaGeo learns to stop guessing wildly and starts aiming for the bullseye, specifically getting better at understanding tricky phrases like "past the bakery, then turn left."

Why is this a big deal?

It handles the messy human stuff: Humans are bad at giving perfect addresses. We say "the place with the blue door" or "near the big tree." Traditional systems fail here. ReaGeo understands the vibe of the location.
It's one step, not ten: The old way was a relay race with many runners (and many chances to drop the baton). ReaGeo is a sprinter who runs the whole distance in one go.
It can draw shapes: It's not just for single points. If you ask for "the whole shopping district," ReaGeo can predict a whole cloud of points that covers that area, like drawing a shape on a map.

The Bottom Line

ReaGeo turns the difficult math of finding a location into a simple language game. By teaching a smart AI to "think" about geography and "practice" until it gets the distance right, it can find your location even if your description is vague, messy, or full of local slang. It's the difference between asking a robot to read a map and asking a local friend to give you directions.

1. Problem Statement

Geocoding is the process of converting human-readable location descriptions (e.g., addresses, relative directions) into machine-readable geographic coordinates.

Limitations of Traditional Methods: Current state-of-the-art approaches rely on multi-stage pipelines involving Named Entity Recognition (NER), hierarchical parsing, and vector/text retrieval against structured geographic databases. These methods suffer from:
- Workflow Complexity: Multiple stages increase latency and engineering overhead.
- Error Propagation: Errors in early stages (e.g., entity extraction) cascade through the pipeline.
- Fragility: They struggle with ambiguous, incomplete, or colloquial queries (e.g., "50 meters north of the intersection") and semantic inconsistencies (dialects, abbreviations) due to rigid reliance on database matching.
Goal: To develop a unified, end-to-end framework that directly maps text queries to geographic coordinates, leveraging the reasoning capabilities of Large Language Models (LLMs) to handle complex spatial relationships without external retrieval databases.

2. Methodology: ReaGeo Framework

ReaGeo reformulates geocoding as a text generation task rather than a retrieval task. It converts geographic coordinates into Geohash sequences (Base32 encoded strings), allowing an LLM to autoregressively generate coordinates.

Core Architecture

Backbone: Uses Qwen2.5-3B, a decoder-only LLM, chosen for its balance of reasoning capability, inference speed, and resource efficiency.
Input/Output:
- Input: Natural language address or relative location query.
- Output: A Geohash string (length 9, offering ~2.4m precision), which is decoded into latitude/longitude coordinates.
Training Stages:
1. Supervised Fine-Tuning (SFT) with Chain-of-Thought (CoT):
  - Mechanism: The model is trained to output intermediate reasoning steps enclosed in <thinking> tags before generating the final Geohash.
  - Purpose: To explicitly model the association between toponyms and formal addresses. For ambiguous queries, the model reasons about neighboring addresses and relative positions to compensate for missing input information.
  - Inference: The thinking tags can be disabled to reduce latency while retaining the learned reasoning patterns.
2. Reinforcement Learning (RL) with GRPO:
  - Algorithm: Uses Group Relative Policy Optimization (GRPO) instead of PPO. GRPO calculates advantages relative to a group of candidate outputs for the same input, eliminating the need for a separate value model and improving training stability.
  - Reward Function: A novel distance-deviation-based reward is designed:
    $R(y_{pred}, y_{true}) = \sqrt{T - \frac{D(y_{pred}, y_{true})}{S}}$
    Where $D$ is the geodesic distance (WGS-84), $T=100$ , and $S=1000$ . This rewards the model for minimizing the physical distance between the predicted and ground-truth coordinates.
  - Data Augmentation: Training data includes "Anchor Offset" samples (e.g., "200m North of X") to teach the model precise directional offsets.

3. Key Contributions

First End-to-End LLM Geocoder: Proposes ReaGeo, the first framework to directly map text to Geohash sequences, eliminating the need for complex NER, entity linking, and vector retrieval pipelines.
Reasoning-Enhanced Training: Integrates Chain-of-Thought (CoT) to improve the model's ability to handle vague relative location descriptions and infer spatial relationships from context.
Specialized RL Optimization: Develops a distance-deviation reward function within a GRPO framework, specifically tailored to optimize spatial accuracy and resolve implicit location descriptions.
Versatility: Demonstrates the ability to predict not just single points, but also non-point geometric regions (line-shaped roads and polygonal administrative areas) by leveraging beam search to generate multiple candidate coordinates.

4. Experimental Results

The model was evaluated on a dataset of ~240k samples from Beijing (covering both rural and urban areas), split into "Base Data" (standard addresses) and "Anchor Offset Data" (relative location descriptions).

Performance Metrics:
- ADD (Average Deviation Distance): Lower is better.
- Acc@k: Percentage of predictions within $k$ meters (100, 200, 500m).
Comparisons: ReaGeo was compared against traditional methods (NER+Levenshtein, Vector Retrieval), commercial APIs (Baidu Maps, Tencent Maps), and raw LLMs (Qwen3-Max).
Key Findings:
- Superior Accuracy: ReaGeo achieved the lowest ADD and highest Acc@k scores in both rural and urban settings.
  - Urban Base Data: ADD of 119.6m (vs. 162.6m for the next best Vector-Retrieval method) and 97.2% accuracy within 500m.
  - Anchor Offset Data: Significantly outperformed all baselines, with an ADD of 208.7m (vs. 378.1m for the next best) and 96.3% accuracy within 500m.
- Robustness: While traditional methods and raw LLMs degraded significantly on relative location queries (Anchor Offset), ReaGeo maintained high precision due to its RL training on offset descriptions.
- Generalization: The model showed strong generalization to unseen intercardinal directions (e.g., Northeast) and successfully visualized non-point geometries (roads and districts) via heat maps.
- Ablation: Removing CoT or GRPO resulted in significant performance drops and increased error rates, confirming the necessity of both reasoning and RL components.

5. Significance and Future Work

Significance: ReaGeo represents a paradigm shift in geocoding from retrieval-based to generative-based approaches. It proves that LLMs can learn the joint distribution of text and geography, handling ambiguity and complex spatial reasoning better than rigid database lookups. This simplifies system architecture and improves robustness for real-world, messy user queries.
Limitations:
- Currently limited to a single city (Beijing); scaling to a global scale requires larger models and more computational resources.
- The distinction between "memorization" of training data and "reasoning" about new locations remains an open research question.
Future Directions: Plans include scaling the model to hundreds of cities, optimizing the trade-off between model size and geographic coverage, and integrating visual data (multi-modal geocoding) to further enhance accuracy.