🤖 machine learning

From Symbolic to Geometric: Enabling Spatial Reasoning in Large Language Models

This paper introduces the Spatial Language Model (SLM), a novel multimodal framework that replaces symbolic pattern matching with native geometric reasoning by treating spatial locations as a first-class modality, thereby significantly outperforming existing LLMs on spatial reasoning tasks through the use of a dedicated training dataset and a new evaluation benchmark.

Original authors: Chen Chu, Bita Azarijoo, Li Xiong, Khurram Shafique, Cyrus Shahabi

Published 2026-06-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chen Chu, Bita Azarijoo, Li Xiong, Khurram Shafique, Cyrus Shahabi

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

The Problem: The "Map-Reading" Robot vs. The "Real-World" Navigator

Imagine you have a very smart robot (a Large Language Model, or LLM) that has read every book, map description, and travel blog in the world. If you ask it, "Is Paris in France?" it will answer correctly because it has memorized that fact from its training data.

However, this paper argues that the robot is actually bad at real spatial reasoning. It's like a student who has memorized the answers to a math test but doesn't actually understand how to do the math.

The Current Flaw (Symbolic Reasoning): When you ask the robot, "How far is it from Point A to Point B?" or "Do these two shapes overlap?", it tries to guess based on words it has seen before. It's like trying to calculate the distance between two cities by guessing based on the names of the cities rather than looking at a map.
The Result: The robot often hallucinates (makes things up), gets confused by new locations it hasn't seen in books, or fails at tasks requiring precise geometry (like measuring the area of a weirdly shaped polygon). It treats space as a list of words, not as a physical reality.

The Solution: The "Spatial Language Model" (SLM)

The authors created a new type of AI called SLM (Spatial Language Model). Think of this as giving the robot a pair of glasses that let it see geometry directly, rather than just reading descriptions of it.

Instead of feeding the robot text like "The park is at 34.05° N, 118.24° W," the SLM feeds the robot a mathematical "fingerprint" (a vector) that represents the actual shape and location of the park.

The Analogy:

Old Way (Symbolic): You describe a friend to a security guard by saying, "He is tall, wears a red hat, and lives on 5th Street." The guard has to guess who you mean based on a list of descriptions.
New Way (Geometric/SLM): You hand the guard a photo of the friend. The guard doesn't need to guess; they can see exactly who it is and where they are standing relative to others.

How They Built It

To teach the robot this new way of seeing, the researchers had to build three new things:

A New Language (Interleaved Grounding): They invented a special way to talk to the robot. Instead of just writing "Long Beach," they wrote <NAME> <PHRASE> <GEO>. The <GEO> part is where they plug in the mathematical "fingerprint" of Long Beach. This forces the robot to look at the shape and location, not just the name.
A New Textbook (Spatial Instruction Dataset): There were no existing textbooks that taught robots how to do geometry using these mathematical fingerprints. So, the authors wrote a new "textbook" with 30,000 practice problems. They broke complex questions down into small, logical steps (like a "chain of thought") so the robot could learn the process of measuring and comparing, not just the answers.
A New Test (SpatialEval): They built a new exam to test if the robot actually learned. Unlike old tests that just asked trivia, this test gives the robot raw coordinates and asks it to calculate distances, check if shapes touch, or find the closest object.

What Happened? (The Results)

The researchers ran a series of tests comparing the new SLM against top-tier robots (like GPT-4 and Llama) that were forced to use the old "word-only" method.

Accuracy: The old robots were terrible at geometry. They would guess wildly when asked to measure distances or areas. The new SLM was significantly more accurate, often getting the math right because it was actually "seeing" the geometry.
Generalization: The old robots failed when asked about places they hadn't read about in books. The new SLM could handle new, unseen locations because it understood the rules of space, not just the names of places.
Speed & Efficiency: The old robots often tried to "think" out loud, writing pages of text to do simple math, which took a long time and used a lot of computer power. The new SLM did the math internally and instantly, using far fewer computer resources.

The Bottom Line

The paper concludes that to make AI truly smart about the physical world, we can't just feed it more books. We have to teach it to understand space as geometry.

By treating location as a "first-class citizen" (a primary type of data, just like text or images) rather than just a word in a sentence, the new SLM can reason about the world the way humans do: by understanding shapes, distances, and positions directly, rather than just guessing based on vocabulary.

Technical Summary: From Symbolic to Geometric: Enabling Spatial Reasoning in Large Language Models

Problem Statement

Current Large Language Models (LLMs) exhibit apparent spatial reasoning capabilities, but these are largely symbolic rather than geometric. LLMs rely on pattern matching over discrete linguistic tokens (e.g., place names, coordinate strings) learned from text corpora, rather than performing explicit geometric computations over continuous spatial representations. This reliance on symbolic reasoning leads to several critical limitations:

Brittleness: Models struggle with structured spatial inference (e.g., topological reasoning, metric consistency) and often produce contradictory or logically invalid conclusions when faced with tasks requiring precise numerical operations.
Generalization Failure: Reasoning is constrained to entities explicitly encoded during training; models fail to generalize to unseen locations or new spatial configurations.
Computational Inefficiency: Current approaches often depend on external tools (agentic frameworks) or complex prompt engineering to invoke map APIs or search engines. This introduces latency, high token consumption, and a reliance on external knowledge sources rather than intrinsic model capability.
Lack of Native Support: LLMs lack native support for continuous spatial representations, explicit geometric computation, and structured spatial operators.

Methodology

The authors propose the Spatial Language Model (SLM), the first multimodal LLM designed to treat geospatial information as a first-class modality, enabling intrinsic geometric spatial reasoning within the inference process.

1. Unified Geometric Representation

Instead of encoding spatial relationships implicitly through text, SLM maps heterogeneous geospatial entities (points, polylines, polygons, regions) into a shared embedding space.

Encoder: The model utilizes Geo2Vec as a unified spatial encoder to convert raw geospatial entities into vector representations ( $V_E$ ) that capture both geometric shape and absolute location.
Adapter: A spatial adapter ( $f_{ada}$ ) projects these geometric vectors into the LLM's pretrained word embedding space, allowing the language model to process spatial data natively alongside text.

2. Interleaved Geometric Grounding

To handle spatial entities appearing in natural language, the authors introduce an interleaved prompt format: <NAME> <PHRASE> <GEO>.

During inference, the <GEO> token is replaced by the learned spatial representation vector ( $V'_E$ ) of the specific entity.
This allows the model to perform reasoning directly on the geometric embeddings rather than retrieving knowledge associated with textual place names.
For training, the model is presented with index tokens (e.g., <E_0>) to distinguish entities, while pretraining/fine-tuning aligns actual place names with their corresponding spatial locations.

3. Spatial Instruction Dataset Construction

To address the lack of training data aligning spatial representations with reasoning, the authors constructed a Spatial Instruction Dataset (30k queries).

Structure: Complex queries are decomposed into a geospatial chain-of-thought format using atomic geometric operations (e.g., Area(<NAME>), Distance(<NAME_0>, <NAME_1>)).
Categories: The dataset covers three types of questions:
1. Single Geographic Entity: Attribute extraction (e.g., area, length).
2. Spatial Relationship: Metric and topological relations between entities (e.g., distance, containment).
3. Spatial Qualifier: Comparative computations among multiple entities.
Prompt Evolution: To ensure linguistic diversity and prevent overfitting, seed questions from experts were expanded using LLMs (Llama-3.3-70B, ChatGPT) to generate varied phrasings while maintaining the same spatial reasoning intent.
Agent-Guided Answer Construction: An LLM agent was used to generate structured, minimal, and logically consistent reasoning traces for complex queries, ensuring the model learns compositional reasoning patterns rather than unstructured hallucinations.

4. SpatialEval Benchmark

The authors introduced SpatialEval, a new benchmark designed to evaluate both symbolic and geometric spatial reasoning.

Design: Unlike previous benchmarks relying solely on text, SpatialEval explicitly includes the actual geographic coordinates or vector embeddings of entities mentioned in queries.
Scope: It evaluates performance across metric reasoning (distance), geometric reasoning (area), and topological reasoning (classification) using diverse input modalities (names, coordinates, embeddings).

Key Contributions

SLM Architecture: The first geospatial multimodal LLM that integrates intrinsic geometric spatial reasoning by treating spatial representations as a primary modality.
Spatial Instruction Dataset: The first dataset aligning spatial representations, atomic geometric operations, and natural language instructions, supporting diverse input types.
SpatialEval Benchmark: A comprehensive evaluation framework designed to assess LLMs' geometric spatial reasoning capabilities under heterogeneous spatial inputs.
Unified Representation Learning: A method to compress complex symbolic representations of geospatial entities into single vector representations, enabling efficient reasoning across heterogeneous entity types.

Experimental Results

Experiments were conducted on three datasets (Beijing, Los Angeles, US Boundaries) covering POIs, road networks, and building footprints.

Performance vs. Baselines: SLM significantly outperformed existing LLM-based approaches (including Qwen3, DeepSeek-R1, Llama-3.3, and proprietary models like Gemini and ChatGPT) that rely on symbolic reasoning or external tools.
- Accuracy: SLM achieved the lowest Mean Absolute Error (MAE) in area and length estimation and the highest accuracy in topological and nearest-neighbor tasks. For instance, in distance estimation, SLM (MAE ~784.9) vastly outperformed symbolic baselines (e.g., Qwen3-8B at ~7816.0).
- Validity: SLM maintained a 100% response validity rate across all tasks, whereas reasoning-oriented models like DeepSeek often failed to produce valid answers due to computational errors or token limits.
Generalization:
- Zero-Shot Transfer: SLM demonstrated strong generalization when transferring between cities (e.g., Los Angeles to Beijing), with minimal performance degradation in relational tasks, unlike symbolic models which struggled significantly.
- Query Complexity: SLM generalized to queries involving 6–7 entities (unseen during training), maintaining stable accuracy.
Efficiency:
- Token Consumption: SLM required a constant, minimal number of input/output tokens (~30), compared to thousands for symbolic reasoning methods and high overhead for agentic approaches.
- Latency: Inference time for SLM was comparable to symbolic methods without reasoning traces (~1.9s) but with significantly higher accuracy.
Convergence: During fine-tuning, the geometric SLM showed a steady decline in error, whereas symbolic fine-tuning (using text-based coordinates) exhibited instability and failed to converge on complex tasks like distance estimation.

Significance and Claims

The paper claims that the fundamental gap between symbolic reasoning over spatial language and genuine geometric reasoning over space has been bridged by SLM.

Intrinsic Capability: By integrating geometric representations directly into the model, SLM achieves robust spatial reasoning without relying on external tools or agentic pipelines.
Efficiency and Scalability: The approach offers a more efficient and scalable solution for spatial intelligence, reducing token consumption and latency while improving accuracy.
Future Direction: The authors posit that intrinsic spatial understanding is a fundamental yet underexplored capability for future LLMs, offering a complementary direction to agentic approaches. They suggest this capability may help LLMs better organize their semantic space and improve performance on general GeoQA tasks.

The work emphasizes that while symbolic reasoning is brittle and limited by training data, geometric reasoning via unified representations allows models to generalize to unseen entities and perform precise spatial computations natively.