Task-Oriented Spatial Representations for LLM-Based Spatial Reasoning: An Empirical Framework With Multi-Layer Controls
IEEE Access · 2026 · 14 · 69161-69182
DOI: 10.1109/access.2026.3690723 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Chae-Seok Lee und 3 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Institute of Information and Communications Technology Planning and Evaluation, Korean Government (MSIT), Development of Next-Generation AI Integrated Mobility Simulation and Prediction/Application Technologies for Metropolitan Cities.
Abstract (Originalsprache)
Large language models (LLMs) have demonstrated considerable capability in spatial reasoning, yet prior work has focused primarily on diagnosing the limits of LLM spatial cognition rather than identifying which input representations improve performance. This study addresses a complementary question: how should spatial information be structured to support LLM-based spatial reasoning? We propose a five-level spatial abstraction hierarchy—from raw coordinates (L1) through semantic labels (L2), natural-language relations (L3a), domain-specific rules (L3b), to pre-computed distance tables (L4)—and evaluate its effects on four task families: category-filtered retrieval (H1), relative position reasoning (H2), rule-constrained inference (H3), and distance comparison (H4). Experiments span controlled synthetic urban scenes (8, 12, and 16 objects), four real-world geographic information system (GIS) scenes derived from OpenStreetMap (Daejeon, South Korea), four open-source models, and two commercial models (GPT-5.2, Claude Sonnet 4.5), with paired bootstrap confidence intervals, McNemar tests, and five-seed replication. The results comprise one design precondition validation and three principal findings. As a precondition, semantic labels are confirmed as a structural prerequisite for category-based tasks—an expected but empirically verified requirement ( $p \lt 0.001$ ). The three findings are: 1) explicit natural-language relation sentences interfere with coordinate-based position reasoning ( $-19.7 \pm 7.0{\,}{\%\text {p}}$ ), with transition analysis confirming that 83% of answer flips are correct-to-incorrect and conflict-trust experiments revealing a wrong-follow rate of 25.6%, although this interference is model- and context-dependent; 2) domain-specific rule injection provides conditional benefit modulated by scene complexity ( $+18.3 \pm 4.9{\,}{\%\text {p}}$ , with substantial scene-level variation); and 3) distance tables yield gains inversely related to the model’s numerical computation ability (+ 2.7 to $+84.1{\,}{\%\text {p}}$ ). These results establish that effective AI-ready spatial representations should be optimized for task relevance and model fit—prioritizing semantic grounding and task-contingent augmentation over information density.
Förderung
- Institute of Information and Communications Technology Planning and Evaluation
- Korean Government (MSIT), Development of Next-Generation AI Integrated Mobility Simulation and Prediction/Application Technologies for Metropolitan Cities
Zitieren
APA
Lee, C., Park, D., Choi, J., & Chang, H. (2026). Task-Oriented Spatial Representations for LLM-Based Spatial Reasoning: An Empirical Framework With Multi-Layer Controls. IEEE Access, 14, 69161-69182. https://doi.org/10.1109/access.2026.3690723
Vancouver
Lee C, Park D, Choi J, Chang H. Task-Oriented Spatial Representations for LLM-Based Spatial Reasoning: An Empirical Framework With Multi-Layer Controls. IEEE Access. 2026;14:69161-69182. doi:10.1109/access.2026.3690723
BibTeX
@article{Lee2026,
author = {Lee, Chae-Seok and Park, Dae-Seung and Choi, Jae-Min and Chang, Ho-Jong},
title = {Task-Oriented Spatial Representations for LLM-Based Spatial Reasoning: An Empirical Framework With Multi-Layer Controls},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {69161--69182},
doi = {10.1109/access.2026.3690723}
}
EUVIMED