Bridging Textual Logic and Geometric Structures

130 concepts110 papers1 proposal

Why do state-of-the-art large language models (LLMs) effortlessly pass professional exams and generate complex code, yet stumble when asked to determine if two geometric shapes overlap or calculate a simple bearing between two GPS coordinates?

The root of this deficiency lies in a fundamental representation mismatch: language models are trained on one-dimensional sequences of text, whereas the physical world operates in multi-dimensional, continuous space.

The Dimensionality Mismatch

As Ankit Goyal et al. point out in Think Visually: Question Answering through Virtual Imagery, traditional language models represent spatial information through word embeddings or symbolic logic. These linear representations are structurally unsuited for geometric reasoning. To resolve spatial dependencies using text alone, a model must execute a cascade of explicit logical deductions. In contrast, a structured visual format makes these same relationships immediately apparent without intensive computational search. Text is a secondary representation of a primary spatial reality; forcing a model to perform coordinate transformations or estimate relative positions using only vector manipulations of tokens is mathematically inefficient.

Why Coordinates and Code Fail to Bridge the Gap

A natural counterargument is that spatial data can simply be tokenized. We can feed models raw numbers, such as 3D coordinates or GPS points. However, empirical evidence shows that scaling parameters or injecting raw numbers does not grant models spatial common sense.

Thinh Hung Truong et al. demonstrate this bottleneck in GPSBench: Do Large Language Models Understand GPS Coordinates?, revealing a persistent performance gap between a model's ability to retrieve semantic geographic facts and its ability to perform geometric operations on raw coordinates. Similarly, Yansong Ning et al. find in UrbanKGent: A Unified Large Language Model Agent Framework for Urban Knowledge Graph Construction that text-based inference routinely fails at calculating structural connectivity or overlap from raw coordinates. The model processes the numbers as text strings rather than points in a continuous space.

To bypass this, models often attempt to externalize spatial reasoning by generating procedural code or vector graphics (like SVGs). Yet, this approach merely shifts the problem. The authors of GeoGramBench: Benchmarking the Geometric Program Reasoning in Modern LLMs show that while models can parse and generate procedural drawing syntax, they consistently fail to synthesize the global spatial relationships described by that code.

As observed by Yejin Bang et al. in A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivity, code generation allows models to approximate layouts, but lacks the geometric precision required to maintain exact spatial proportions. The model understands the local recipe for the image, but not the global spatial reality the recipe produces.

Internalizing Space via Virtual Imagery and Latent Constructions

To overcome this spatial blindness, recent architectures attempt to give models an internal "mind's eye." Instead of forcing spatial reality into a 1D text string, these methods map linguistic instructions into structured internal representations that mirror physical layouts.

Ankit Goyal et al. pioneered this with the Dynamic Spatial Memory Network (DSMN), which tracks spatial layouts by generating and reasoning over internal virtual images. By shifting the problem from symbolic logic to visual memory, the model bypasses the need for complex coordinate arithmetic.

A particularly elegant evolution of this idea addresses the challenge of auxiliary constructions in geometry—the act of drawing a new line or point to reveal a hidden mathematical relationship. Traditional models cannot dynamically alter an input image or text description to add these elements.

Haiying Xu et al. resolve this in LatentGeo: Learnable Auxiliary Constructions in Latent Space for Multimodal Geometric Reasoning by allowing Multimodal LLMs to project and manipulate these missing geometric elements directly within a continuous latent space. By internalizing constructions as hidden mathematical vectors rather than discrete tokens or raw pixels, the model bridges the gap between symbolic proofs and continuous spatial layouts.

[Textual Problem] ---> [Traditional LLM] ---------> [1D Token Search (Fails)]
       |
       v
[LatentGeo Approach] -> [Continuous Latent Space] -> [Auxiliary Construction Vectors] -> [Accurate Geometric Proof]

Structural Decomposition and Reverse Verification

Beyond architectural changes, improving spatial reasoning requires restructuring how models decompose geometric problems. Spyridon Mouselinos et al. note in Beyond Lines and Circles: Unveiling the Geometric Reasoning Gap in Large Language Models that models trained on linear text fail to maintain a consistent internal representation of a visual layout during multi-step constructions, leading to spatial hallucinations.

To counteract this, Xiaofeng Wang et al. introduce GeoCoT in Do Large Language Models Truly Understand Geometric Structures?, a method that forces the model to explicitly break down complex geometric structures into isolated relational components before attempting a proof. This prevents the model from arriving at correct answers through superficial text-matching or coincidental guesses.

Complementing this, Linger Deng et al. develop a reverse chain-of-thought framework in R-CoT: Reverse Chain-of-Thought Problem Generation for Geometric Reasoning in Large Multimodal Models that generates theorem-grounded diagrams and back-tracks the logical steps. By forcing the model to generate the spatial layout from the theorem backward, it builds a bi-directional mapping between text logic and geometric space, ensuring that the final reasoning chain is anchored to a physically valid structure.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_spatial_reasoning_gap_fc7a83c8,
  title = {spatial reasoning gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatial-reasoning-gap/rdr_fc7a83c849b543d1bf15f078d9a4d6bb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_spatial_reasoning_gap_fc7a83c8_v1,
  title = {spatial reasoning gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatial-reasoning-gap/rdr_fc7a83c849b543d1bf15f078d9a4d6bb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake