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

Abstract: The image geolocalization task aims to predict the location where an image was taken anywhere on Earth using visual clues. Existing large vision-LLM (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.

• The paper introduces a novel agentic framework that augments LVLMs with map-API tools to iteratively refine geolocation hypotheses.
• It employs reinforcement learning with a distance-sensitive reward and parallel test-time scaling to boost pass@1 and pass@K accuracy.
• The method is validated on challenging benchmarks like MAPBench, achieving significant improvements over prior models.

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

Introduction

"Thinking with Map: Reinforced Parallel Map-Augmented Agent for Geolocalization" (2601.05432) introduces a map-augmented agentic paradigm for global-scale image geolocalization using Large Vision-LLMs (LVLMs). This approach augments LVLM reasoning with structured interactions over map APIs, explicitly modeling the process as an agent-in-the-map loop. The framework proposes a two-stage optimization protocol: agentic reinforcement learning (RL) post-training to improve tool-use efficacy and a parallel test-time scaling (TTS) method with a trajectory verifier for efficient and robust candidate selection. The work further establishes MAPBench, a challenging, up-to-date geolocalization benchmark emphasizing both easy and hard samples from real-world street-view data. The proposed agent achieves strong empirical results, advancing pass@1 and pass@K accuracy across multiple public and proprietary baselines.

Methodology: Agent-in-the-Map Loop and Map-Augmented Tooling

Problem Reformulation and Tool Suite

The geolocalization task is recast from a conventional discriminative or retrieval approach to an agentic, multi-step reasoning process grounded in explicit map interaction. Rather than relying exclusively on parametric world knowledge, the agent iteratively forms, verifies, and updates location hypotheses through a repertoire of map-based tools—e.g., POI search, detail lookups, static/satellite image queries—mirroring the way humans leverage map software for place recognition. This externalizes much of the knowledge validation step, attenuating hallucinations and biases prevalent in autoregressive LVLMs.

The policy model, $\pi_\theta$, communicates with an API-abstracted map environment, $P_{\text{env}}$. Each agent step involves proposing candidate locations, querying the map environment, and using returned observations to iteratively refine a pool of plausible location candidates, $\mathcal{C}_t$. The process continues either until a convergence criterion is met or the inference budget is exhausted.

Figure 1: Panel (a) shows the agent-in-the-map loop for hypotheses maintenance; (b) illustrates RL training; (c) depicts the parallel test-time scaling and verifier pipeline.

Agentic Reinforcement Learning

Building atop the intrinsic tool-use capabilities of Qwen3-VL, the agent undergoes direct RL post-training using Group Relative Policy Optimization (GRPO). The reward function is distance-sensitive, assigning denser rewards for finer localization (e.g., full reward for ≤500m error) and providing coarse granularity for larger misses. This enables discriminative learning even under ambiguity and supplies an effective backbone for reinforcement across tool-use actions. Critically, the RL phase optimizes pass@N performance and increases sampling efficiency by incentivizing exploration of diverse solution paths rather than memorization.

Parallel Test-Time Scaling with Verifier

A principal insight is that map-augmented reasoning, due to the explicit factual content acquired from APIs, leaves self-contained, verifiable trajectories. Thus, at inference, parallel reasoning rollouts can be generated to explore a broad hypothesis landscape efficiently. A dedicated LVLM-based verifier model evaluates and selects the most plausible prediction given the synthesized evidence across all rollouts. This transfers performance gains from pass@K (oracle among K tried solutions) to pass@1 (top single answer), enabling robust inference even under high uncertainty or weak visual cues.

Figure 2: Example Thinking with Map trajectories from parallel sampling, highlighting causal validation via map-API outputs.

Figure 3: Comparison of parallel sampling strategies, illustrating gains in pass@K as the number of trajectories increases.

Evaluation: MAPBench and Benchmark Performance

Dataset Construction and Difficulty Segmentation

The MAPBench benchmark specifically targets real-world, up-to-date geolocalization by collecting 5,000 Chinese street-view images centered on unique POIs, stratified into easy and hard splits based on zero-shot predictions of large models. This design effectively isolates memorization-based success (easy) from situations requiring deep reasoning and tool-use generalization (hard). Additional evaluations are carried out on IMAGEO-Bench and GeoBench to confirm robustness and cross-dataset transfer.

Quantitative Results

The Thinking with Map agent achieves dominant results on MAPBench-test-hard, improving Acc@500m from 4.02% (Gemini-3-Pro with Google Map/Search) to 14.86%—a relative gain of nearly 11 points. Across benchmarks and granularity thresholds, the method consistently advances over both open-source and commercial closed-source LVLM baselines.

Strong numerical performances are demonstrated across:

• MAPBench-test-easy: 44.98% Acc@500m (vs. 20.86% for Gemini-3-Pro)
• GeoBench: 57.94% Acc@500m (parallel$\times$4 verifier, best in class)
• IMAGEO-2-test: 20.53% Acc@500m (top open-source result for 30B model class)

  Figure 4: Pass@K accuracy curves throughout RL training stages on MAPBench.

The experiments further underline that fine-level localization accuracy is highly sensitive to effective map-tool integration, while coarse-level accuracy remains tied to base model priors. Notably, direct addition of map tools without RL learning can degrade performance at coarse scales, necessitating careful reward shaping and stability mechanisms.

Figure 5: Reward distribution dynamics during RL, showing dense learning signal allocation across spatial thresholds.

Analytical and Ablation Investigations

Ablations dissected the contributions of different tool types and verifier models:

• POI and static map tool usage contributed orders-of-magnitude more to fine localization than image-zoom or generic web search.
• Increasing the number of parallel rollouts escalates pass@K, with diminishing verifier error beyond N=4.
• RL training compresses the variance in prediction accuracy, maximizing the returns from parallel test-time sampling.
• Verifier strength only marginally impacts accuracy for small N; at higher N values, larger verifier models provide a measurable benefit.

Implications, Limitations, and Prospects

The presented work constitutes a significant methodological advance in agentic multimodal reasoning, specifically targeting the class of tasks where world knowledge can and should be externalized through structured tool interfaces. The integration of RL with tool-augmented environments aligns with trends in agentic retrieval-augmented generation (RAG) and interactive tool-use RL observed in recent LLM and LVLM agent literature, including agentic RAG frameworks (Singh et al., 15 Jan 2025), tool RL benchmarks [Qwen3 Technical Report, (Yang et al., 14 May 2025)], and test-time scaling/verification strategies [Snell et al., (Snell et al., 2024)].

Practically, this approach enables robust, up-to-date geolocalization without dependence on massive parametric pretraining or regionally biased retrieval galleries. Theorizing more generally, it exemplifies the power of agent-environment interactivity as an inductive bias—yielding solutions that are less brittle, more interpretable, and natively equipped for real-world operational dynamics.

Nonetheless, the approach is not without limitations:

• Long-horizon, compositional spatial reasoning (e.g., inferring orientation or exploiting relative geometry) remains below human levels.
• RL sample efficiency is currently constrained by the limited, largely Chinese-centric dataset coverage.
• Parallel test-time scaling serves as a pragmatic workaround for model limitations in deep credit assignment and long-chain reflection.

Future developments may involve scaling to global, multi-provider map APIs; transferring this paradigm to broader geospatial and navigation domains; or adopting advanced credit assignment mechanisms as in recent multi-turn agent RL frameworks [Dong et al., (Dong et al., 26 Jul 2025); Feng et al., (Feng et al., 16 May 2025)].

Conclusion

Thinking with Map substantiates a new agentic LVLM paradigm for global image geolocalization. By synchronizing internal reasoning with explicit, factual external validation over map APIs, and employing RL fine-tuning as well as parallel, verifiable test-time sampling, the approach achieves best-in-class fine-grained localization on contemporary benchmarks. The insights extend to broader multimodal agent research, motivating structured, tool-augmented, and verifier-oriented learning for robust real-world AI systems.