Research map / Vision-language and multimodal / Multimodal LLMs
GeoSolver: Scaling Test-Time Reasoning in Remote Sensing with Fine-Grained Process Supervision
Sun Lang, Ronghao Fu, Zhuoran Duan, Haoran Liu, Xueyan Liu, Bo Yang
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 average PRM process score as reward (ablation) | Vanilla GRPO · avg of 6 RS tasks | average score: 67.09 -> 58.13 | Hurt |
| Add drop-moment (process-aware) penalty (ablation) | Vanilla GRPO · avg of 6 RS tasks | average score: 67.09 -> 68.16 | Improved |
| Add entropy-guided tree exploration (ablation) | Vanilla GRPO · avg of 6 RS tasks | average score: 67.09 -> 68.13 | Improved |
| Combine drop-moment penalty with tree exploration (ablation) | Process-Aware Tree-GRPO · avg of 6 RS tasks | average score: 67.09 -> 70.51 | Improved |
| Add SFT on Geo-CoT380k (ablation) | GLM-4.1V-Base · avg of 6 RS tasks | average score: 22.56 -> 60.34 | Improved |
| Remove MCTS-derived samples from GeoPRM training data (ablation) | GeoPRM (GeoSolver, BoN N=32) · object-level | score: 74.26 -> 64.27 | Hurt |
| Remove synthetic hallucination injection from GeoPRM training data (ablation) | GeoPRM (GeoSolver, BoN N=32) · object-level | score: 74.26 -> 70.26 | Hurt |
| Remove synthetic hallucination injection (ablation) | GeoPRM (GeoSolver, BoN N=32) · scene-level N=4 | score: 74.03 -> 74.24 | No clear effect |
| Use lower drop threshold rho=0.2 instead of 0.3 (ablation) | Drop-moment penalty · avg score | average score: 70.51 -> 69.06 | Hurt |
| Use severe penalty gamma=0.3 instead of 0.7 (ablation) | Drop-moment penalty · avg score | average score: 70.51 -> 68.21 | Hurt |
| Use GeoPRM verification (BoN or beam search) vs greedy decoding | GeoSolver · six RS tasks | performance | Improved |
| Use generic VLM or math PRM verifiers instead of GeoPRM | verifier choice · RS tasks | performance | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- EarthDial — EarthDial: Turning Multi-sensory Earth Observations to Interactive Dialogues (CVPR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.92)
- RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow (AAAI 2026 · similarity 0.91)
- GeoCoT: Towards Reliable Remote Sensing Reasoning with Manifold Perspective (CVPR 2026 · similarity 0.90)
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.89)
- Text Before Vision: Staged Knowledge Injection Matters for Agentic RLVR in Ultra-High-Resolution Remote Sensing Understanding (ICML 2026 · similarity 0.89)
- GeoViS: Geospatially Rewarded Visual Search for Remote Sensing Visual Grounding (CVPR 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.
