Research map / Vision-language and multimodal / CLIP and zero-shot
Unlocking Zero-Shot Geospatial Reasoning via Indirect Rewards
Chenhui Xu, Fuxun Yu, Michael J. Bianco, Jacob Kovarskiy, Raphael Tang, Qi Zhang, Zirui Xu, Will LeVine, Brandon Dubbs, Heming Liao, Cassandra Burgess, Suvam Bag and 6 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add indirect-reward RL (cross-view pairing) after scaffolding SFT (ablation) | Qwen2.5-VL-7B · CV-Cities cross-view pairing (in-distribution) | accuracy: 23.1% -> 82.4% | Improved |
| train with SFT on positive examples only instead of RL (ablation) | Qwen2.5-VL-7B · CV-Cities cross-view pairing (in-distribution) | accuracy: 19.0% -> 23.1% | No clear effect |
| add scaffolding SFT before RL (ablation) | Geo-R1-Zero · CV-Cities cross-view pairing (in-distribution) | accuracy: 78.1% -> 82.4% | Improved |
| add scaffolding SFT before RL (ablation) | Geo-R1-Zero · CV-Cities cross-view pairing (in-distribution) | completion length: 587.4 -> 378.8 | Improved |
| SFT alone causes content duplication (ablation) | Geo-SFT · CV-Cities cross-view pairing (in-distribution) | completion length: 204.6 -> 1127.6 | Hurt |
| apply Geo-R1 indirect-reward RL | Qwen2.5-VL-7B · IMAGEO-GSS | city accuracy: 0.277 -> 0.327 | Improved |
| apply Geo-R1 to a different backbone | InternVL3-8B · IMAGEO-GSS | city accuracy: 0.146 -> 0.313 | Improved |
| apply Geo-R1 to InternVL3-8B on the in-distribution pairing task | InternVL3-8B · CV-Cities cross-view pairing (in-distribution) | accuracy: 20.3% -> 84.7% | Improved |
| indirect reward instead of task-direct RL reward (GLOBE-7B) | Geo-R1 vs GLOBE-7B · MP16-Reason | accuracy at distance thresholds: 17.98 vs 17.99 (1km); 93.56 vs 92.52 (2500km) | No clear effect |
| Geo-R1 vs GeoCLIP task-specific model | Geo-R1 · IMAGEO-GSS | median distance (km): 266.8987 -> 69.4008 | Improved |
| add SFT scaffolding vs RL on general-purpose benchmarks (ablation) | Geo-SFT vs Geo-R1 · MEGA-Bench, GPQA, MMMU | accuracy | Hurt |
Closest papers by meaning
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.91)
- Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.90)
- Textual Supervision Enhances Geospatial Representations in Vision-Language Models (ICML 2026 · similarity 0.89)
- RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow (AAAI 2026 · similarity 0.89)
- Recognition through Reasoning: Reinforcing Image Geo-localization with Large Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- Text Before Vision: Staged Knowledge Injection Matters for Agentic RLVR in Ultra-High-Resolution Remote Sensing Understanding (ICML 2026 · similarity 0.88)
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.
