Research map / Vision-language and multimodal / Multimodal LLMs

GeoAgent: Learning to Geolocate Everywhere with Reinforced Geographic Characteristics

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces GeoAgent, a Qwen2.5-VL-7B model trained with human-annotated chain-of-thought SFT followed by GRPO using a geo-similarity reward (spatial + semantic) and a consistency reward from a separate consistency agent, plus the GeoSeek dataset.

Problem

RL-based image geolocation relies on AI-generated chain-of-thought data and exact-text reward matching, which ignores that different descriptions can refer to the same place and may yield inconsistent reasoning.

Approach

Replace exact-match rewards with a geo-similarity reward (distance-based spatial reward plus embedding-based hierarchical semantic reward), and add a consistency reward where a separate LLM agent must infer the answer from the reasoning alone, using human-annotated CoT and a stratified global sampling strategy.

Key results

• On IM2GPS3K, GeoAgent reaches 40.75% city, 58.57% region, 76.21% country and 89.90% continent accuracy, exceeding the listed baselines at region, country and continent levels.
• On GeoSeek-Val, GeoAgent achieves 60.37% country accuracy vs 56.13% for GeoCLIP and 11.13% for base Qwen2.5-VL-7B, with GeoScore 3314.1.
• Ablation on GeoSeek-Val: the directly-judge reward gives 50.81 country accuracy versus 60.37 for geo-similarity reward.

Surprising findings

• Adding the consistency reward alone to SFT-only models slightly decreases city and region accuracy in the ablation.
• The InternVL3-8B + SFT variant performs worse than its zero-shot counterpart on city/region metrics relative to expectations, suggesting base-model choice matters.

Limitations

• Data sourced from player communities focuses on street-view imagery and performs poorly indoors or in scenes with few geographic cues.
• Consistency reward and semantic reward rely on external components (OpenCage geocoding, GPT-4o-based locatability labels, Qwen3-32B consistency agent).

Experimental setup

Datasets: GeoSeek-CoT GeoSeek-Loc GeoSeek-Val OSV-5M GeoComp IM2GPS3K GeoGuessr community data TuXun community data

Benchmarks: IM2GPS3K GeoSeek-Val

Metrics: accuracy at 25km/200km/750km/2500km GeoScore

Baselines: GeoDecoder GeoCLIP Translocator PIGEOTTO G3 Qwen2.5-VL-32B Gemma3-27B InternVL3-14B GeoReasoner GaGA GRE-Suite-SFT GRE-Suite

Builds on: GRPO Qwen2.5-VL LoRA GRE GLOBE

Closest papers by meaning

• Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.89)
• MapEval: A Map-Based Evaluation of Geo-Spatial Reasoning in Foundation Models (ICML 2025 · similarity 0.88)
• GeoMMBench and GeoMMAgent: Toward Expert-Level Multimodal Intelligence in Geoscience and Remote Sensing (CVPR 2026 · similarity 0.87)
• RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow (AAAI 2026 · similarity 0.87)
• GeoReasoner: Geo-localization with Reasoning in Street Views using a Large Vision-Language Model (ICML 2024 · similarity 0.86)
• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.85)