Research map / Vision-language and multimodal / Multimodal LLMs

GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization

NeurIPS 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.

Proposes GeoRouter, an LVLM-based router that chooses per query between retrieval (GeoRanker) and generation (Gemini-2.5-Flash) predictions for worldwide image geolocalization, trained with a distance-aware soft-label objective (DisPO) on a new 100K-sample GeoRouting dataset.

Problem

Worldwide image geolocalization methods are split between retrieval and LVLM generation, which have complementary error profiles; no single paradigm dominates, so a method is needed to pick the better paradigm per query.

Approach

Treat paradigm selection as learned routing: an LVLM reads both paradigms' predictions and retrieved candidates and outputs a scalar score, trained with soft labels derived from the log geodesic error gap between the two paradigms (DisPO) instead of hard binary labels.

Key results

• GeoRouter reaches 61.61% average accuracy on IM2GPS3K versus 59.24% for pure generation and the best listed baseline GeoRanker (about 60.8% mean from its row), with 20.82% at 1km and 50.48% at 25km.
• On YFCC4K, GeoRouter improves the 25km accuracy from 43.54% (GeoRanker) to 46.01%, and reaches 58.29% average versus 56.18% for pure retrieval.
• Ablation on IM2GPS3K shows replacing DisPO with standard BCE lowers the mean from 61.60 to 60.66.

Surprising findings

• Qwen2-VL-2B routing already outperforms the strongest single-paradigm baseline GeoRanker, per the backbone-scale appendix.
• Oracle routing still leaves a gap (61.61% vs 65.84% average on IM2GPS3K), indicating headroom in policy optimization.

Limitations

• Routing relies on pre-computed predictions from both paradigms, so inference requires running a commercial API (Gemini) plus retrieval.
• Gap to the oracle upper bound remains substantial.
• Reported gains on YFCC4K at 1km are negligible (0.12% relative).

Experimental setup

Datasets: MP16-Pro IM2GPS3K YFCC4K

Benchmarks: IM2GPS3K YFCC4K

Metrics: geolocalization accuracy at 1km, 25km, 200km, 750km, 2500km routing accuracy

Baselines: [L]kNN PlaNet CPlaNet ISNs Translocator GeoDecoder GeoCLIP Img2Loc PIGEON G3 GeoToken GeoRanker

Builds on: GeoRanker Img2Loc LoRA AdamW

Closest papers by meaning

• GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.92)
• Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.90)
• G3: An Effective and Adaptive Framework for Worldwide Geolocalization Using Large Multi-Modality Models (NeurIPS 2024 · similarity 0.89)
• Around the World in 80 Timesteps: A Generative Approach to Global Visual Geolocation (CVPR 2025 · similarity 0.89)
• PIGEON: Predicting Image Geolocations (CVPR 2024 · similarity 0.88)
• GeoReasoner: Geo-localization with Reasoning in Street Views using a Large Vision-Language Model (ICML 2024 · similarity 0.87)