Structured Maps for Landmark-Based Route Instructions with LLMs

Mifu Matsumoto; Yuki Nakanishi; Tomoya Hashiguchi; Hiroaki Ohshima

cs.CLOct 27, 2026

Structured Maps for Landmark-Based Route Instructions with LLMs

Authors: Mifu Matsumoto, Yuki Nakanishi, Tomoya Hashiguchi, Hiroaki Ohshima

Organizations: University of Hyogo, Hyogo, Japan

Abstract

To generate landmark-based route instructions, models must understand both map geometry and human landmark preferences. Although large language models (LLMs) show great abilities in language tasks, the best way to structure map data for these models is not yet established. This paper proposes a structured text representation for landmark selection tasks in route navigation, utilizing SVG format to encode spatial features. The structured text maintains a geometric connection with rendered map images, allowing language models to directly access spatial and semantic data. Using a new dataset created from human-written route descriptions and OpenStreetMap data, we evaluate models on two tasks: landmark selection at turns and around destinations. Experiments with five LLMs show that the proposed structured text input performs as well as or better than image-based PNG inputs. In addition, task-specific fine-tuning enables smaller local LLMs to effectively leverage map layouts to achieve high selection accuracy. These results demonstrate that structured text is an efficient and lightweight interface for spatial reasoning in human-oriented navigation.

Figures & tables

Figure 1: Overview of the landmark selection task. The framework takes local geographic data as input, which is then processed by an LLM to select and output the optimal landmark.

Figure 2: Examples of dataset instances constructed from raw access descriptions. The left panel shows an instance for Task 1, where an intersection, an underlined landmark, and a turning direction are extracted. The right panel shows an instance for Task 2, where an underlined landmark around the destination is identified.

Figure 3: Example of a local map represented in the normalized $(100\times 100)$ coordinate space, originally expressed in Japanese.

	Task 1		Task 2
Model	Batch size	Epochs	Batch size	Epochs
Llama-3-ELYZA-JP-8B	128	95	128	52
ELYZA-Thinking-1.0-Qwen-32B	16	21	4	10
Gemma3:27B	4	6	2	7

Table 1: Fine-tuning batch sizes and optimal checkpoint epochs for each model and task.

	Task 1		Task 2
Model	PNG	SVG	PNG	SVG
GPT-5.2	34%	34%	40%	55%
Gemma3:27B	14%	27%	26%	39%
Llama4	21%	25%	24%	28%
Llama-3-ELYZA-JP-8B	N/A	25%	N/A	35%
ELYZA-Thinking-1.0-Qwen-32B	N/A	24%	N/A	39%

Table 2: Landmark selection accuracy comparison between PNG image input and structured SVG text input.

	Task 1		Task 2
Model	Base	Fine-Tuned	Base	Fine-Tuned
Llama-3-ELYZA-JP-8B	24%	49%	35%	48%
ELYZA-Thinking-1.0-Qwen-32B	24%	38%	39%	54%
Gemma3:27B	27%	53%	39%	36%

Table 3: Landmark selection accuracy before and after fine-tuning across both tasks.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: Prompt templates for Task 1.

Panel of Figure 6 — Figure 6: Prompt templates for Task 2.

Spatial Reasoning in Language Models

See all 52 papers

The Sword, Shield, and Achilles' Heel: Characterizing the Linguistic Inductive Bias of Large Language Models for Spatial Reasoning in Navigation Planning
May 29, 2026Xudong Zhang, Jian Yang, Shengkai Wang +5Spatial Reasoning in Language Models Large Language Model-Based Robot Planning
1East China Normal University, China · 2Information Engineering University, China · 3Zhengzhou University, China
NaViRrator: Robot Navigation from Human-Readable Maps through a Learned Visual Route
Sep 18, 2026Ayun Lee, Jiseon Kim, Giseop KimRobot Navigation Visual Robot Navigation
Department of Robotics and Mechatronics Engineering, DGIST, Daegu, Republic of Korea
MapReason-OSM: Can Vision-Language Models Make Graph-Verifiable Mobility Decisions from Street Maps ?
Jun 21, 2026Srinivas Venkatanarayanan, Clement Pakkam IsaacGeospatial Reasoning Spatial Reasoning Benchmarks
NVIDIA · University of Central Florida · University of South Florida

May 29, 2026J/K move · Enter open · S save

The Sword, Shield, and Achilles' Heel: Characterizing the Linguistic Inductive Bias of Large Language Models for Spatial Reasoning in Navigation Planning

Xudong Zhang, Jian Yang, Shengkai Wang, Jiangpeng Tian+4

1East China Normal University, China · 2Information Engineering University, China · 3Zhengzhou University, China

Spatial Reasoning in Language Models Large Language Model-Based Robot Planning Causal Interventions in Language Models Spatial Reasoning

Open paper AskPDF arXiv

Large Language Model (LLM)-based navigation systems commonly construct explicit spatial representations (e.g., topological graphs, semantic raster maps) and translate them into textual descriptions as LLMs' inputs. However, the linguistic structures of such text-based spatial representations and the choices of contextual features (e.g., topology, geometry) they contain are often treated as neutral engineering decisions rather than key factors that shape LLMs' behavior. To fill the gap, we propose a dual-interventional framework that disentangles linguistic structures from different contextual cues to evaluate the linguistic inductive bias of LLMs for navigation planning. In the framework, representation intervention varies the linguistic format and the degree of linguistic compression, clarifying when linguistic representations support or inhibit navigation planning. Context intervention, combined with contextual feature combination and conflict probing, explicitly clarifies the preferences and weaknesses of LLMs when processing different contextual cues. Experiments across diverse spatial reasoning tasks and multiple model scales reveal a consistent pattern: topological information is a sturdy shield and the backbone of robust planning; linguistic format is a double-edged sword whose effect depends on model size, task demands, and the compression level; and semantic information is a fatal Achilles' heel -- incorrect semantic cues can systematically derail the planning process. Overall, our study shows that effective text-based spatial representations in LLM-based navigation should preserve topological integrity, calibrate representational compression to model capacity, and ensure semantic correctness, rather than simply adopting a single representation. Our code is publicly available at https://github.com/jonesdong150/LLM-Navigation-Inductive-Bias.

Abstract

Spatial Reasoning in Language Models

Explore similar work

The Sword, Shield, and Achilles' Heel: Characterizing the Linguistic Inductive Bias of Large Language Models for Spatial Reasoning in Navigation Planning

NaViRrator: Robot Navigation from Human-Readable Maps through a Learned Visual Route

MapReason-OSM: Can Vision-Language Models Make Graph-Verifiable Mobility Decisions from Street Maps ?