Title: Bridging the Gap: Transitioning Large Language Models from Symbolic to Geometric Spatial Reasoning

Abstract:

While contemporary large language models (LLMs) frequently demonstrate an apparent aptitude for spatial reasoning, this proficiency is predominantly \emph{symbolic} in nature. It stems from pattern recognition within spatial linguistic data rather than genuine \emph{geometric} manipulation of space. Due to their reliance on discrete tokens, LLMs inherently lack the capacity for continuous spatial representations, explicit geometric calculations, and structured spatial operators.

To overcome these constraints, we present the \emph{Spatial Language Model (SLM)}, pioneering the integration of location data as a primary modality within multimodal LLMs to facilitate geometric spatial reasoning during inference. Unlike traditional approaches that depend on textual descriptions, SLM functions directly on learned spatial representations.

To ensure robust training, we have developed a \emph{Spatial Instruction Dataset} that harmonizes spatial representations, fundamental geometric operations, and natural language commands. Additionally, we introduce \emph{SpatialEval}, a novel benchmark specifically engineered to assess spatial reasoning capabilities across four key dimensions: attributes, distance, topology, and relative positioning.

Comprehensive experiments reveal that SLM substantially surpasses current LLM-based methods that depend on symbolic reasoning techniques, such as prompt engineering or textual abstraction. These results underscore the advantages of embedding geometric spatial representations to achieve reliable spatial reasoning. The source code for model training, the evaluation benchmark, the instruction dataset, and model checkpoints are publicly accessible at: \hyperlink{https://github.com/chuchen2017/SLM}{https://github.com/chuchen2017/SLM}.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC