PaperlayerView exact revision on arXiv

Source-linked AI summary

Thinking with Map: Reinforced Parallel Map-Augmented Agent for Geolocalization

Yuxiang Ji, Yong Wang, Ziyu Ma, Yiming Hu, Hailang Huang, Xuecai Hu, Guanhua Chen, Liaoni Wu, Xiangxiang Chu

arXiv:2601.05432v1cs.CVcs.AIcs.CL

TL;DR

LVLM geolocalization methods largely overlook maps, which people use to verify location hypotheses. This paper equips an LVLM with map tools and combines agentic reinforcement learning with parallel test-time scaling. It reports better performance than open- and closed-source models on most metrics, including fine-localization gains on evaluated benchmarks.

  • Problem

    Existing LVLM geolocalization methods rely on internal reasoning, while maps—used to verify location hypotheses—are almost absent from those methods.

    Source: §1 · p. 2

  • Method

    The paper equips an LVLM with map tools for iterative hypothesis verification, then combines agentic reinforcement learning with parallel verifier-based test-time scaling.

    Source: §1 · pp. 2–3

  • Results

    Across four test sets, the method achieves the best performance against open- and closed-source models on most metrics; MAPBench-test-hard Acc@500m rises from 4.02% to 14.86% over Gemini-3-Pro.

    Source: §5.2 · p. 7

  • Takeaways & Limitations

    The combined optimization gains pass@N capability within a single query.

    Source: §6 · p. 9

    The authors report that map-use ability remains far below human performance and that limited agentic RL training data constrain learning in open environments.

    Source: Limitation · p. 9

Abstract

from arXiv · show

The image geolocalization task aims to predict the location where an image was taken anywhere on Earth using visual clues. Existing large vision-language model (LVLM) approaches leverage world knowledge, chain-of-thought reasoning, and agentic capabilities, but overlook a common strategy used by humans -- using maps. In this work, we first equip the model \textit{Thinking with Map} ability and formulate it as an agent-in-the-map loop. We develop a two-stage optimization scheme for it, including agentic reinforcement learning (RL) followed by parallel test-time scaling (TTS). The RL strengthens the agentic capability of model to improve sampling efficiency, and the parallel TTS enables the model to explore multiple candidate paths before making the final prediction, which is crucial for geolocalization. To evaluate our method on up-to-date and in-the-wild images, we further present MAPBench, a comprehensive geolocalization training and evaluation benchmark composed entirely of real-world images. Experimental results show that our method outperforms existing open- and closed-source models on most metrics, specifically improving Acc@500m from 8.0\% to 22.1\% compared to \textit{Gemini-3-Pro} with Google Search/Map grounded mode.

1 Introduction

Existing LVLM geolocalization methods rely on internal reasoning despite maps being a common way to verify location hypotheses. The paper introduces map interaction as an agentic loop and combines parallel test-time scaling with reinforcement learning, evaluating the approach on MAPBench and other benchmarks.

1 Introduction

  • Existing LVLM geolocalization methods use chain-of-thought and external tools but still depend on internal reasoning over visual evidence.The task requires inspecting clues such as climate, architecture, and cultural context and reasoning across them.Source: §1 · p. 2
  • The agent uses map tools in a multi-turn loop, iteratively proposing location hypotheses and verifying visual clues against structured map information.The tools include keyword search, POI details lookup, and static map queries.Source: §1 · p. 2
  • Parallel test-time scaling lets the model explore multiple hypotheses, while agentic reinforcement learning targets pass@K and more effective sampling.The paper motivates parallel exploration by the context growth and marginal gains associated with longer sequential reasoning.Source: §1 · pp. 2–3
  • The method outperforms open- and closed-source models on most metrics across MAPBench and other evaluated benchmarks.MAPBench covers up-to-date Chinese urban street-view and POI images, with easy and hard cases.Source: §1 · p. 3
  • MAPBench includes up-to-date, broadly covered Chinese urban street-view and POI images, divided into easy and hard cases by clue distinctiveness.Hard cases have less distinctive clues and are described as unlikely to be encountered during pre-training.Source: §1 · p. 3

2 Related Work

Related work spans retrieval- and classification-based geolocalization, direct LVLM inference, and interactive agents using reinforcement learning or test-time scaling.

2 Related Work

  • Conventional approaches use retrieval over geo-tagged galleries or classification into geocells, while recent LVLM methods infer locations directly from images.The cited LVLM approaches do not use a database or map partitioning (Jia et al., 2025; Li et al., 2025a; Wang et al., 2024a; Li et al., 2025b; Huang et al., 2025).Source: §2 · p. 3
  • Although explicit reasoning can reduce LVLMs’ black-box nature, it cannot prevent their hallucinations and biases.Source: §2 · p. 3
  • Agentic reinforcement learning has been used to improve tool use and long-horizon decision-making, with GeoVista applying the approach to geolocalization.Prior work also explores parallel sampling, sequential revision, and multi-agent exploration for test-time scaling.Source: §2 · p. 3
  • Some studies argue that general search tools offer very limited benefits for localization.Source: §2 · p. 3

3 Method

Thinking with Map equips an LVLM with map tools and structures geolocalization as an agent-in-the-map loop. Agentic RL improves sampling, followed by parallel test-time scaling that selects among candidate paths.

3.1 Thinking with Map

  • Thinking with Map iteratively proposes or verifies location hypotheses, using map observations and interaction history to update an implicit candidate pool.The model selects a final answer from the pool when sufficiently confident or when its interaction budget is exhausted.Source: §3.1 · p. 4
  • The method provides POI search, static and satellite maps, and image zoom to retrieve location details and cross-check visual clues.Two map API providers support global geolocalization because map-service availability varies by region.Source: §3.1 · p. 4

3.2 RL for Map-augmented Agent

  • Agentic RL uses GRPO to optimize map-interacting trajectories, with structured answers and distance-based rewards providing a verifiable learning signal.The reward scheme reflects localization granularity, including fine-level 500m and city-level 25km ranges.Source: §3.2 · pp. 4–5

3.3 Parallel Test-time Scaling

  • Verifier@N closely matches oracle best@N, with almost negligible verifier loss at N = 2 or 4.Parallel sampling lets the model explore multiple Thinking with Map hypotheses and aggregate self-verifiable trajectories, transferring gains from pass@K to pass@1.Source: §3.3 · p. 5

4 Dataset

The section motivates MAPBench by identifying limitations in existing geolocalization benchmarks, then describes its construction and difficulty-tiered test set.

4 Dataset

  • Existing benchmarks are often outdated, skew toward landmark memorization, and lack coverage of China, motivating a more current and broadly covered evaluation set.Obsolete POIs may conflict with current map or web information, while landmark-style images may test memorization rather than reasoning and external-knowledge use.Source: §4 · pp. 5–6
  • MAPBench contains 5,000 nearby street-view images centered on distinct POIs and is split into 2,500 training and 2,500 test samples.Its images are described as up-to-date and broadly covering China.Source: §4 · p. 6
  • The test set labels 599 samples easy and 1,901 hard using agreement among three base models on predictions within 10km of ground truth.The easy split evaluates base-model memorization and world knowledge; the hard split assesses agentic capabilities.Source: §4 · p. 6

5 Experiment

The experiments compare Thinking with Map against open- and closed-source models across multiple benchmarks and localization distances. They report gains from map tools, reinforcement learning, and parallel test-time scaling, alongside differences in verifier performance.

5.1 Experimental Setup

  • Evaluation compares the method, built on Qwen3-VL-30B-A3B-Instruct, with closed- and open-source models on MAPBench, IMAGEO-Bench, and GeoBench.RL training uses MAPBench training data and 2,000 IMAGEO-2 examples.Source: §5.1 · pp. 6–7

5.2 Main Results

  • Acc@500m improves from 37.79% to 57.94% on GeoBench and from 16.33% to 20.53% on IMAGEO-2-test; the method is best on most metrics across four test sets.On MAPBench-test-hard, Gemini-3-Pro's Acc@500m rises from 4.02% to 14.86% with Thinking with Map.Source: Table 4 · p. 7, §5.2 · p. 7

5.1 Experimental Setup

  • Accuracy is measured at six distance thresholds, from Acc@500m to Acc@750km, with a prediction correct when it falls within the corresponding threshold.The thresholds are 500m, 2km, 10km, 25km, 200km, and 750km.Source: §5.1 · p. 7

5.3 Quantitative Analysis

  • Figure 5 tracks pass@K during RL training: accuracy variance falls, best@N improves above 10km, while Best@500m shows little gain and can limit exploration.The passage attributes the lower variance to shrinking Range@2/4.Source: Figure 5 · p. 8, §5.3 · p. 8
  • At N=2, verifier choice has little impact and a 30B model is sufficient; as parallel size increases, verification gets harder and model capacity matters more.Table 6 reports the verifier-model ablation.Source: Table 6 · p. 8, §5.3 · pp. 8–9

5.2 Main Results

  • Thinking with Map already outperforms other approaches before parallel test-time scaling, which then brings further gains that grow with the number of parallel samples.The reported improvement trend is consistent with the base model using parallel test-time scaling.Source: §5.2 · p. 8

5.3 Quantitative Analysis

  • Map, image-zoom, and web-search tools improve localization below 2km but hurt it above 200km; map_tool raises Acc@500m from 1.12% to 16.16%.Image zoom and web search yield only marginal improvements.Source: §5.3 · p. 8

6 Conclusion

The conclusion presents map-augmented geolocalization as an agent-in-the-map loop and combines agentic RL with parallel test-time scaling. The method outperforms open- and closed-source models on most metrics.

6 Conclusion

  • The agent-in-the-map loop proposes location hypotheses, retrieves map information, cross-validates evidence, and converges on a decision.Source: §6 · p. 9
  • The two-stage optimization combines agentic RL and parallel test-time scaling to gain pass@N capability within one query.Source: §6 · p. 9
  • The method outperforms open- and closed-source models on most metrics.Source: §6 · p. 9

Limitation

The authors describe limits in map-use ability, RL training data, and single-agent long-horizon problem solving. They frame parallel test-time scaling as an interim solution.

Limitation

  • The model's map-use ability remains below human performance, and it does not infer orientation from relative spatial relationships.Source: Limitation · p. 9
  • Very limited agentic RL training data constrain learning in open environments.Source: Limitation · p. 9
  • Parallel test-time scaling is an interim workaround; building a single agent with stronger long-horizon problem-solving remains open.Source: Limitation · p. 9

A Datasets

MAPBench samples real-world images around points of interest in Chinese cities, while IMAGEO-2 and GeoBench supply additional test data. The appendix also lists RL training and parallel-testing hyperparameters.

A Datasets

  • MAPBench contains 5,000 images sampled from 5,000 valid POIs across 20 Chinese cities, with a street-view or storefront photo selected for each POI.The authors describe the construction as up-to-date and broadly covering.Source: A Datasets · p. 12
  • The authors include IMAGEO-2 and GeoBench alongside MAPBench for worldwide coverage and timely image sources.Source: A Datasets · p. 12
  • Table 7 lists hyperparameters for Thinking with Map RL training and parallel testing.Source: Table 7 · p. 12
  • IMAGEO-2 is a subset of IMAGEO-Bench, constructed from crowdsourced Google Map POI images; data released by Yan et al. (2023) were compiled and filtered to a final 2,929 images.Source: A Datasets · p. 12
  • GeoBench is used for testing and contains 512 normal photos, 512 panoramas, and 108 satellite images.The images come from Internet sources, the Mapillary API, and Sentinel-2 Level-2A imagery, respectively.Source: A Datasets · p. 12

B.1 Implementation Details

This section outlines the implementation setup and prompt conventions for Thinking with Map, alongside pointers to training-dynamics and ablation visuals.

B.1 Implementation Details

  • Agentic RL used the VeRL codebase, and RL training and other experiments ran on 32 NVIDIA H20 GPUs.The specific RL-training and parallel-testing hyperparameters are listed in Table 7.Source: §B.1 · p. 12
  • Thinking with Map and base-model prompts use fixed JSON outputs, but only Thinking with Map adds tool-use guidance.The verifier receives the original query and multiple parallel trajectories, then returns an answer in the same JSON format.Source: §B.1 · pp. 12–13
  • The appendix identifies Figure 6 as RL reward dynamics, Table 8 as verifier-model ablations, and Table 9 as RL-algorithm ablations.Table 8 covers GeoBench and IMAGEO; Table 9 covers RL algorithms.Source: Figure 6 · p. 13, Table 8 · p. 13, Table 9 · p. 13

B.2 Training Dynamics of RL

The reported training reward rises overall, while oscillating toward stability in the latter half of training.

B.2 Training Dynamics of RL

  • 0.25 to 0.45: training reward rose from the early stage to the end, which the authors say demonstrates a positive effect on localization accuracy.During the second epoch, the reward gradually oscillated toward stability, which the authors say suggests more data may be needed.Source: §B.2 · p. 13

B.3 Ablation Study on RL Algorithm

The ablations report that vanilla GRPO performs better than the tested pass@K-oriented RL algorithms, while verifier models can sometimes produce answers beyond the sampled candidates.

B.3 Ablation Study on RL Algorithm

  • Pass@K-GRPO and PKPO perform substantially worse than vanilla GRPO on this task, so the authors use GRPO-trained models for parallel TTS.Table 9 reports the RL-algorithm ablation.Source: §B.3 · p. 13

B.4 More Ablation Studies on Verifier Models

  • On GeoBench and IMAGEO-2-test, a different-base-model verifier such as GPT-5 can outperform Best@N (Oracle) in a few cases.The authors suggest that in these cases the verifier identifies more plausible answers along a Thinking with Map trajectory, rather than only selecting among existing candidates.Source: §B.4 · p. 13
Loading 2601.05432v1…