Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MapTrace: Scalable Data Generation for Route Tracing on Maps

Artemis Panagopoulou, Aveek Purohit, Achin Kulshrestha, Soroosh Yazdani, Mohit Goyal

TL;DR

This work tackles the limited spatial reasoning of multimodal language models by introducing MapTrace, a task that demands pixel-precise route tracing on maps. It presents a scalable synthetic data pipeline that generates 23k path annotations across 4k maps, enabling effective fine-tuning of open and proprietary MLLMs. Experiments show substantial gains in path-tracing robustness and alignment metrics on real-world MapBench data, underscoring that fine-grained spatial reasoning can be taught with synthetic supervision. The approach opens the door to scalable spatial benchmarks and applications in robotics, AR/VR, and wayfinding systems, while highlighting ongoing domain-transfer challenges from synthetic to real maps.

Abstract

While Multimodal Large Language Models have achieved human-like performance on many visual and textual reasoning tasks, their proficiency in fine-grained spatial understanding, such as route tracing on maps remains limited. Unlike humans, who can quickly learn to parse and navigate maps, current models often fail to respect fundamental path constraints, in part due to the prohibitive cost and difficulty of collecting large-scale, pixel-accurate path annotations. To address this, we introduce a scalable synthetic data generation pipeline that leverages synthetic map images and pixel-level parsing to automatically produce precise annotations for this challenging task. Using this pipeline, we construct a fine-tuning dataset of 23k path samples across 4k maps, enabling models to acquire more human-like spatial capabilities. Using this dataset, we fine-tune both open-source and proprietary MLLMs. Results on MapBench show that finetuning substantially improves robustness, raising success rates by up to 6.4 points, while also reducing path-tracing error (NDTW). These gains highlight that fine-grained spatial reasoning, absent in pretrained models, can be explicitly taught with synthetic supervision.

MapTrace: Scalable Data Generation for Route Tracing on Maps

TL;DR

This work tackles the limited spatial reasoning of multimodal language models by introducing MapTrace, a task that demands pixel-precise route tracing on maps. It presents a scalable synthetic data pipeline that generates 23k path annotations across 4k maps, enabling effective fine-tuning of open and proprietary MLLMs. Experiments show substantial gains in path-tracing robustness and alignment metrics on real-world MapBench data, underscoring that fine-grained spatial reasoning can be taught with synthetic supervision. The approach opens the door to scalable spatial benchmarks and applications in robotics, AR/VR, and wayfinding systems, while highlighting ongoing domain-transfer challenges from synthetic to real maps.

Abstract

While Multimodal Large Language Models have achieved human-like performance on many visual and textual reasoning tasks, their proficiency in fine-grained spatial understanding, such as route tracing on maps remains limited. Unlike humans, who can quickly learn to parse and navigate maps, current models often fail to respect fundamental path constraints, in part due to the prohibitive cost and difficulty of collecting large-scale, pixel-accurate path annotations. To address this, we introduce a scalable synthetic data generation pipeline that leverages synthetic map images and pixel-level parsing to automatically produce precise annotations for this challenging task. Using this pipeline, we construct a fine-tuning dataset of 23k path samples across 4k maps, enabling models to acquire more human-like spatial capabilities. Using this dataset, we fine-tune both open-source and proprietary MLLMs. Results on MapBench show that finetuning substantially improves robustness, raising success rates by up to 6.4 points, while also reducing path-tracing error (NDTW). These gains highlight that fine-grained spatial reasoning, absent in pretrained models, can be explicitly taught with synthetic supervision.
Paper Structure (31 sections, 1 equation, 5 figures, 4 tables)

This paper contains 31 sections, 1 equation, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Novel Task: MapTrace
  4. Task Definition
  5. Evaluation Metrics
  6. Evaluation Metrics
  7. MapTrace: Synthetic Data Generation
  8. Experimental Setup
  9. Results
  10. Conclusion
  11. Limitations and Future Work
  12. Synthetic Data Generation Details
  13. Map Categories
  14. Prompt Templates
  15. Map Description Generator.
  16. ...and 16 more sections

Figures (5)

  • Figure 1: MapTrace: Given a start and end location, the model outputs a valid path that respects map constraints
  • Figure 2: MapTrace Synthetic data generation pipeline. A LLM generates various map descriptions, which are rendered into images by a text-to-image model. Candidate path masks are extracted via dominant-color selection and filtered by a Mask Critic. Valid masks are converted into a pixel-graph to compute shortest-path candidates, which are then judged by a Path Critic for quality and traversability.
  • Figure 3: Distribution of map types in the synthetic MapTrace dataset.
  • Figure 4: Qualitative examples of MapTrace paths on diverse map types.
  • Figure 5: Qualitative examples comparing the fine-tuned Gemini-2.5-Flash (red) to the base model (blue). The fine-tuned model adheres more closely to the intended routes and avoids non-traversable regions.