Research map / Vision-language and multimodal / Multimodal LLMs
GRE Suite: Geo-localization Inference via Fine-Tuned Vision-Language Models and Enhanced Reasoning Chains
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 the GRE Suite (GRE30K dataset, GRE reasoning model built on Qwen2.5-VL-7B with CoT SFT and two-stage GRPO, and GREval-Bench) for reasoning-augmented worldwide image geo-localization.
Problem
Worldwide image geo-localization lacks robust, explainable reasoning and open-ended coordinate prediction; existing alignment-based and grid/retrieval methods depend on candidate databases and offer no reasoning-quality evaluation.
Approach
Cold-start SFT on o3-generated, human-refined chain-of-thought geo-localization data, followed by two-stage rule-based GRPO with distance-based threshold rewards (and yes/no judgment rewards), applied to a VLM to produce open-ended coordinates with explicit reasoning.
Key results
- On Im2GPS3k, GRE (CI+I+II) reaches 11.33% at 1km and 69.33% at 750km, improving over the prior SOTA by +0.5 and +1.7 points at those thresholds respectively.
- On GWS15k, GRE reaches 78.28% at 2500km, +4.2 over the prior SOTA SeekWorld (67.3%) per the paper's reported table.
- On GREval-Bench, GRE attains 59.54 CoT quality and 83.16% at 2500km, outperforming Qwen2.5VL-32B (55.56 CoT quality, 65.00% at 2500km).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GRE instead of GeoCLIP | Worldwide geo-localization · Im2GPS3k | accuracy within 1 km: 10.8 -> 11.3 | Helped |
| Use GRE instead of GeoReasoner | Worldwide geo-localization · GWS15k | accuracy within 750 km: 65.4 -> 54.8 | Hurt |
| Use GRE instead of Qwen2.5VL-32B | GREval-Bench · GREval-Bench | CoT quality: 55.56 -> 59.54 | Helped |
| Add cold-start initialization (CI) (ablation) | Qwen2.5-VL-7B · Im2GPS3k | accuracy within 1 km: 3.20 -> 7.77 | Helped |
| Add RL stage I after cold-start initialization (ablation) | Qwen2.5-VL-7B · Im2GPS3k | accuracy within 1 km: 7.77 -> 7.16 | Hurt |
| Add RL stage II after cold-start initialization (ablation) | Qwen2.5-VL-7B · Im2GPS3k | accuracy within 1 km: 7.77 -> 10.96 | Helped |
| Add RL stages I and II after cold-start initialization (ablation) | Qwen2.5-VL-7B · Im2GPS3k | accuracy within 1 km: 10.96 -> 11.33 | Helped |
| Add cold-start initialization (CI) (ablation) | Qwen2.5-VL-7B · GWS15k | accuracy within 1 km: 0.05 -> 0.45 | Helped |
| Add RL stage I after cold-start initialization (ablation) | Qwen2.5-VL-7B · GWS15k | accuracy within 1 km: 0.45 -> 0.35 | Hurt |
| Add RL stage II after cold-start initialization (ablation) | Qwen2.5-VL-7B · GWS15k | accuracy within 1 km: 0.45 -> 0.88 | Helped |
| Add RL stages I and II after cold-start initialization (ablation) | Qwen2.5-VL-7B · GWS15k | accuracy within 1 km: 0.88 -> 0.91 | Helped |
| Add cold-start initialization (CI) (ablation) | LLaVA-v1.5-7B · Im2GPS3k | accuracy within 1 km: 1.7 -> 4.2 | Helped |
Surprising findings
- Qwen2.5VL-7B base scores 50.36 CoT quality on GREval-Bench, higher than the larger Qwen2.5VL-32B's 55.56 only partly; smaller Qwen models showed weaker implicit cue extraction.
- Stage I RL (judgment-based) alone slightly decreased performance on several granularities on Im2GPS3k, attributed to a mismatch between training and test reward types.
Limitations
- Substantial compute needs (8 NVIDIA H20 GPUs) for training.
- API costs for o3-based dataset generation.
- Stage I RL reward misalignment with test tasks reduced some granularities.
Experimental setup
Datasets: MP-16 Im2GPS3k GWS15k OSV-5M Google StreetView
Benchmarks: Im2GPS3k GWS15k GREval-Bench OSV-5M Google StreetView
Metrics: geolocation accuracy at 1km/25km/200km/750km/2500km threshol CoT quality (recall, RefCLIPScore, BERTScore) average distance
Baselines: GeoCLIP GeoReasoner SeekWorld Translocator GeoDecoder PlaNet ISNs Molmo-D-7B LLaVA-v1.5-7B InternVL3 InternVL2.5 Qwen2.5-VL
Builds on: GRPO DeepSeek-R1 Qwen2.5-VL MP-16
Closest papers by meaning
- Recognition through Reasoning: Reinforcing Image Geo-localization with Large Vision-Language Models (NeurIPS 2025 · similarity 0.94)
- GeoReasoner: Geo-localization with Reasoning in Street Views using a Large Vision-Language Model (ICML 2024 · similarity 0.92)
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.91)
- TimeSpot: Benchmarking Geo-Temporal Understanding in Vision–Language Models in Real-World Settings (ICML 2026 · similarity 0.90)
- ERGeoBench: A Comprehensive Benchmark for Embodied Reasoning and Geo-localization in Multimodal Large Language Models (ICML 2026 · similarity 0.89)
- GTR-Bench: Evaluating Geo-Temporal Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.