Research map / Vision-language and multimodal / Multimodal LLMs

Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach

AAAI 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.

Geo-R trains Qwen2.5-VL-7B for retrieval-free image geolocalization with hierarchical Chain-of-Region reasoning labels derived from GPS coordinates, SFT on 500K samples, then GRPO with a Haversine-distance reward and a hard-subset selection of 200K samples.

Problem

Image geolocalization with vision-language models currently relies on retrieval databases or synthetic reasoning annotations, which limits interpretability and generalization.

Approach

Derive country/region/city reasoning chains by reverse-geocoding ground-truth GPS coordinates (no synthetic teacher labels), then optimize coordinates with GRPO using a Haversine-distance reward gated by a format reward, and select a hard, long-tail training subset to avoid vanishing advantages.

Key results

• Geo-R reaches 18.10% at 1km and 86.42% at 2500km on IM2GPS3K, beating retrieval-free GeoCLIP (14.11% at 1km).
• On YFCC4K, Geo-R achieves 10.47% at 1km and 75.84% at 2500km, best among retrieval-free methods in the table.
• For Qwen2.5-VL-3B, RL on hard 200K samples after SFT-500k raises 1km accuracy from 12.6% to 14.8% on IM2GPS3K.

Surprising findings

• Random-sample RL on the SFT-10k or SFT-500k setting sometimes lowers accuracy at coarse thresholds (e.g., 750km and 2500km), while hard-sample RL does not consistently do so.
• Plain location-only SFT gives 1km gains smaller than CoR gains in the 3B ablation, but the 3B CoT baseline has a 200km value (47.5%) above the CoR value (44.8%).

Limitations

• Ablations run only on the 3B model due to RL compute cost.
• Main table compares against retrieval-based methods and reports only two benchmarks.
• Reward formulation and claims about reward variance and vanishing advantages are not quantitatively analyzed.

Experimental setup

Datasets: MP16-Pro MP16-Rand-500K MP16-Hard-200K IM2GPS3K YFCC4K

Benchmarks: IM2GPS3K YFCC4K

Metrics: geodesic accuracy at 1km, 25km, 200km, 750km, 2500km

Baselines: [L]kNN Img2Loc PIGEON G3 Geo-Ranker PlaNet CPlaNet ISNs Translocator GeoDecoder GeoCLIP GLOBE

Builds on: GRPO Qwen2.5-VL DeepSeek-R1 MP16-Pro

Closest papers by meaning

• Recognition through Reasoning: Reinforcing Image Geo-localization with Large Vision-Language Models (NeurIPS 2025 · similarity 0.93)
• GRE Suite: Geo-localization Inference via Fine-Tuned Vision-Language Models and Enhanced Reasoning Chains (NeurIPS 2025 · similarity 0.91)
• GeoReasoner: Geo-localization with Reasoning in Street Views using a Large Vision-Language Model (ICML 2024 · similarity 0.91)
• GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization (NeurIPS 2026 · similarity 0.90)
• Textual Supervision Enhances Geospatial Representations in Vision-Language Models (ICML 2026 · similarity 0.90)
• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.89)