GeoSolver: Scaling Test-Time Reasoning in Remote Sensing with Fine-Grained Process Supervision

Lang Sun; Ronghao Fu; Zhuoran Duan; Haoran Liu; Xueyan Liu; Bo Yang

cs.CVMar 10, 2026

GeoSolver: Scaling Test-Time Reasoning in Remote Sensing with Fine-Grained Process Supervision

Authors: Lang Sun, Ronghao Fu, Zhuoran Duan, Haoran Liu, Xueyan Liu, Bo Yang

Organizations: College of Computer Science and Technology, Jilin University, Changchun 130012, China · Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education Jilin University

Abstract

While Vision-Language Models (VLMs) have significantly advanced remote sensing interpretation, enabling them to perform complex, step-by-step reasoning remains highly challenging. Recent efforts to introduce Chain-of-Thought (CoT) reasoning to this domain have shown promise, yet ensuring the visual faithfulness of these intermediate steps remains a critical bottleneck. To address this, we introduce GeoSolver, a novel framework that transitions remote sensing reasoning toward verifiable, process-supervised reinforcement learning. We first construct Geo-PRM-2M, a large-scale, token-level process supervision dataset synthesized via entropy-guided Monte Carlo Tree Search (MCTS) and targeted visual hallucination injection. Building upon this dataset, we train GeoPRM, a token-level process reward model (PRM) that provides granular faithfulness feedback. To effectively leverage these verification signals, we propose Process-Aware Tree-GRPO, a reinforcement learning algorithm that integrates tree-structured exploration with a faithfulness-weighted reward mechanism to precisely assign credit to intermediate steps. Extensive experiments demonstrate that our resulting model, GeoSolver-9B, achieves state-of-the-art performance across diverse remote sensing benchmarks. Crucially, GeoPRM unlocks robust Test-Time Scaling (TTS). Serving as a universal geospatial verifier, it seamlessly scales the performance of GeoSolver-9B and directly enhances general-purpose VLMs, highlighting its remarkable cross-model generalization.

Figures & tables

Figure 1: Overall framework of our proposed GeoSolver. To enable reliable geospatial reasoning, we first construct the Geo-PRM-2M dataset via (a) Entropy-Guided MCTS and (b) Synthetic Hallucination Injection. The trained GeoPRM is then seamlessly integrated to (c) scale test-time reasoning via advanced search strategies during inference, and (d) align the base policy via Process-Aware Tree-GRPO during training.

Figure 2: Illustration of the Process-Aware Tree-GRPO. The reasoning tree dynamically expands by identifying high-entropy tokens, followed by trajectory rollouts. Leaf nodes receive outcome rewards which are further refined by GeoPRM’s drop-moment penalty. These process-aware signals are then aggregated upwards to compute Local Advantage ( $LA$ ) and Global Advantage ( $GA$ ) for fine-grained policy optimization.

Method	Visual Grounding				Object Detection		Object Counting
Method	DIOR	RRSIS	VRS	RSVG	DOTA	HRRSD	RSOD	VHR	DOTA	HRRSD
Closed-source Commercial Vision-Language Models
Claude-sonnet-4 [ 1 ]	8.87	11.62	6.65	2.53	3.89	14.87	51.5	25.0	25.17	50.11
Gemini-2.0-flash [ 11 ]	11.40	16.05	13.52	1.75	14.30	28.92	63.5	39.0	29.36	54.65
ChatGPT-5 [ 42 ]	8.27	10.60	5.65	1.49	8.66	13.15	40.0	58.0	36.20	58.50
Open-source Reasoning Vision-Language Models

Table 1: Comparison of GeoSolver with state-of-the-art models on object-level remote sensing tasks. We report mIoU and mAP@50 for Visual Grounding (DIOR-RSVG, RRSIS-D, VRSBench, and RSVG) and Object Detection (DOTAv2-val and HRRSD), alongside Accuracy for Object Counting (RSOD, NWPU-VHR, DOTAv2-val, and HRRSD). The best results are bolded, and the second-best are underlined.

Method	Scene Classification					VQA		Image Caption
Method	NWPU	AID	RS19	SIRI	UCM	VRS	HR	RSIT	NWPU	RSIC
Closed-source Commercial Vision-Language Models
Claude-sonnet-4	58.44	60.33	76.32	64.33	67.86	58.95	55.94	20.14	28.32	11.58
Gemini-2.0-flash	74.89	76.00	90.00	72.00	85.95	61.92	49.95	15.73	20.55	10.85
ChatGPT-5	82.22	75.50	95.53	75.00	88.57	62.20	65.94	27.27	39.62	16.83
Open-source Vision-Language Models

Table 2: Comparison of GeoSolver with state-of-the-art models on scene-level tasks. We report Accuracy for Scene Classification (NWPU-RESISC45, AID, WHU-RS19, SIRI-WHU, and UCMerced) and VQA (VRSBench and RSVQA-HR). For Image Captioning (RSITMD, NWPU-Captions, and RSICD), we report the BLEU-4 metric.

Panel of Figure 4 — Figure 4: Cross-model generalization capabilities of GeoPRM using BoN performance. The gray bars indicate the base model performance, while the red gradient bars show consistent accuracy gains as the generation budget ( $N$ ) increases. Horizontal dashed lines represent the baseline performance of fully fine-tuned remote sensing models. Notably, with $N=32$ , GeoPRM-guided generalists surpass these domain experts.

Models	Tree	PRM		VG	OC	Det	IC	SC	VQA	Avg
Models	Tree	APS	PA	(mIoU)	(Acc)	(mAP)	(BLEU-4)	(Acc)	(Acc)	Avg
Base (GLM-4.1V-Base)	$-$	$-$	$-$	34.48	16.96	3.56	12.43	59.78	8.16	22.56
+ SFT	$-$	$-$	$-$	53.77	49.73	59.42	44.90	86.01	68.23	60.34
Vanilla GRPO	✗	✗	✗	59.31	55.56	68.89	54.89	91.74	72.17	67.09
+ Average Process Score	✗	✓	✗	48.69	48.22	50.76	46.45	87.21	67.44	58.13
+ Process-Aware	✗	✗	✓	60.12	59.21	69.90	54.67	92.01	73.02	68.16

Table 5: Ablation study of different RL alignment strategies on GeoSolver. We evaluate the progressive integration of Vanilla GRPO, Average Process Score (APS), our Process-Aware (PA), and Tree-based exploration.

Models	Dataset	Object-level				Scene-level
Models	Dataset	N $=$ 4	N $=$ 8	N $=$ 16	N $=$ 32	N $=$ 4	N $=$ 8	N $=$ 16	N $=$ 32
GeoSolver	Geo-PRM-2M	68.44	70.92	73.26	74.26	74.03	75.87	76.97	77.63
	w/o $S_{MC}$	62.74	63.60	63.82	64.27	69.53	70.21	69.82	70.77
	w/o $S_{SHI}$	66.28	67.81	69.72	70.26	74.24	75.46	76.37	76.89
Qwen3-VL-8B	Geo-PRM-2M	28.70	30.98	32.56	34.13	48.51	50.45	51.93	52.99
	w/o $S_{MC}$	24.87	25.38	25.22	26.11	45.11	45.88	46.04	45.97

Table 6: Ablation of Geo-PRM-2M training data components evaluated using BoN performance. We compare reward models trained on the full dataset against those trained without intrinsic MCTS samples (w/o

S_{MC}

) or without Synthetic Hallucination Injection (w/o

S_{SHI}

). Performance is evaluated on both GeoSolver and generic model (Qwen3-VL-8B) across scaling budgets (

N

Figure 5: Qualitative example of GeoPRM verification. We visualize two reasoning trajectories for an aircraft counting task. In the top erroneous attempt, GeoSolver hallucinates a non-existent aircraft at Step 5, which GeoPRM instantly penalizes with a sharp score drop (from 0.966 to 0.228). The bottom trajectory demonstrates how GeoSolver avoids this error to reach a fully grounded conclusion.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

$\rho$	$\gamma$	Score
0.3	0.7	70.51
0.3	0.3	68.21
0.3	0.5	69.66
0.3	0.9	70.15
0.2	0.7	69.06
0.4	0.7	69.42

Appendix

Table 7: Ablation study of the drop-moment hyperparameters. We report the average score across multiple tasks.

Task	Dataset	Stage	Split	Samples	QA Examples
Visual Grounding	DIOR-RSVG [ 70 ]	SFT/RL	train/test	31k/8k	User : Where is a slender small train station in the middle? Answer : [280,371,503,420].
	VRSBench [ 27 ]	SFT/RL	train/test	20k/16k
	RRSIS-D [ 33 ]	$-$	test	16k
	RSVG [ 55 ]	$-$	test	1k
Object Detection	DOTAv2 [ 60 , 23 , 13 ]	SFT/RL	train/test	29k/9k	User : Detect all plane in the image. Answer : [[63, 70, 28, 138], [45, 120, 50, 185], [226, 148, 307, 203]].
	HRRSD [ 74 ]	SFT/RL	train/test	11k/11k

Appendix

Table 8: Comprehensive overview of datasets and task formulations. We detail the specific data sources, corresponding alignment stages (SFT/RL), sample sizes across train/test splits, and provide representative instruction-response templates for each of the six core remote sensing tasks evaluated in our study.

Method	VRSBench-VG			DIOR-RSVG			RRSIS-D			RSVG
Method	@0.5	@0.75	mIoU	@0.5	@0.75	mIoU	@0.5	@0.75	mIoU	@0.5	@0.75	mIoU
Close-source Commercial Vision-Language Models
Claude-sonnet-4	4.2	0.6	6.65	4.1	0.0	8.87	5.5	0.5	11.62	1.0	0.0	2.53
Gemini-2.0-flash	8.9	1.6	13.52	3.3	0.8	11.40	10.5	0.0	16.05	0.0	0.0	1.75
ChatGPT-5	2.4	0.2	5.65	5.3	0.0	8.27	8.5	0.5	10.60	0.0	0.0	1.49
Open-source Reasoning Vision-Language Models

Appendix

Table 9: Comparison of GeoSolver with existing generic and RS VLMs on Visual Grounding task.

Method	DOTAv2-val		HRRSD		RSOD		NWPU-VHR
Method	Acc $\uparrow$	MAE $\downarrow$	Acc $\uparrow$	MAE $\downarrow$	Acc $\uparrow$	MAE $\downarrow$	Acc $\uparrow$	MAE $\downarrow$
Close-source Commercial Vision-Language Models
Claude-sonnet-4	25.17	10.232	50.11	2.231	25.0	4.115	51.5	2.205
Gemini-2.0-flash	29.36	15.057	54.65	1.921	39.0	4.095	63.5	1.835
ChatGPT-5	36.20	7.490	58.50	0.787	40.0	1.430	58.0	1.310
Open-source Vision-Language Models

Appendix

Table 10: Comparison of GeoSolver with existing generic and RS VLMs on Object Counting task.

Method	VRSBench-VQA							RSVQA-HR
Method	Category	Existence	Position	Quantity	Scene	Color	Image	Presence	Comp
Close-source Commercial Vision-Language Models
Claude-sonnet-4	43.28	52.78	30.17	66.67	64.79	63.29	91.67	46.95	64.94
Gemini-2.0-flash	44.03	86.11	43.97	46.00	60.56	56.96	95.83	56.94	42.96
ChatGPT-5	39.55	88.89	42.24	47.33	70.42	59.49	87.50	62.94	68.93
Open-source Vision-Language Models

Appendix

Table 11: Comparison of GeoSolver with generic and RS VLMs on Classification and VQA tasks.

Method	RSITMD			NWPU-Captions			RSICD			VRSBench-Cap
Method	B-4	MT	Cr	B-4	MT	Cr	B-4	MT	Cr	B-4	MT	Cr
Close-source Commercial Vision-Language Models
Claude-sonnet-4	20.14	17.15	19.31	28.32	21.98	32.46	11.58	13.90	24.57	14.62	22.36	73.49
Gemini-2.0-flash	15.73	9.27	17.11	20.55	11.42	22.58	10.85	8.71	21.53	14.19	22.30	86.33
ChatGPT-5	27.27	21.10	29.48	39.62	25.69	48.52	16.83	16.73	34.39	18.06	25.11	88.93
Open-source Vision-Language Models

Appendix

Table 12: Comparison of GeoSolver with existing generic and RS VLMs on Image Captioning task.

Panel of Figure 6 — Figure 6: Comprehensive Test-Time Scaling (TTS) curves across all six remote sensing tasks. We compare the scaling behavior of Self-Consistency (Green), Best-of-N (Red), and Beam Search (Blue) as the inference generation budget ( $N$ ) scales from $2^{0}$ to $2^{7}$ . Guided by our token-level GeoPRM, both search strategies unlock robust compute-optimal scaling, significantly and consistently outperforming the unverified Self-Consistency baseline.

Panel of Figure 7 — Figure 7: Comprehensive cross-model generalization evaluation across six remote sensing tasks. Gray bars indicate the greedy decoding performance of the base generalist VLMs (GLM-4.1V-9B, Qwen3VL-8B, Qwen3VL-32B). The red gradient bars demonstrate the progressive performance gains achieved by incorporating GeoPRM as a verifier under Best-of-N ( $N\in\{8,16,32\}$ ) decoding. Horizontal dashed and dotted lines represent the baseline performance of state-of-the-art, fully fine-tuned domain experts (e.g., SkySenseGPT, VHM, EarthDial). Guided by GeoPRM, general-purpose models successfully bridge the domain gap and frequently surpass dedicated remote sensing specialists.

VLM Reasoning

Vision-language model reasoning studies how multimodal systems use visual evidence and language to answer questions, plan, and explain.

See all 846 papers

GeoReason: Bridging Logical Reasoning and Spatial Fidelity in Remote Sensing Segmentation
Oct 16, 2026Cheng Fan, Han Wu, Tingzhang Luo +3Reasoning Segmentation Remote Sensing Image Segmentation
City University of Hong Kong · School of Computer Science, Peking University · Samsung
GeoSearcher: Anchor-Guided Progressive Reasoning for Remote Sensing Visual Grounding with Process Supervision
Jul 1, 2026Dianyu Wang, Peirong Zhang, Xuyang Li +2Vision-Language Grounding VLM Reasoning
Aerospace Information Research Institute, the Key Laboratory of Network Information System Technology (NIST), and the Key Laboratory of Target Cognition and Application Technology (TCAT), Chinese Academy of Sciences, Beijing 100190, China · School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100190, China
Think and Answer ME: Benchmarking and Exploring Multi-Entity Reasoning Grounding in Remote Sensing
Mar 13, 2026Shuchang Lyu, Haiquan Wen, Guangliang Cheng +5Remote Sensing Visual Grounding Visual Grounding
Beihang University, Beijing, China. · University of Liverpool, Liverpool, UK. · Institute of Acoustics, Chinese Academy of Sciences, Beijing, China

Oct 16, 2026J/K move · Enter open · S save

GeoReason: Bridging Logical Reasoning and Spatial Fidelity in Remote Sensing Segmentation

Cheng Fan, Han Wu, Tingzhang Luo, Minjing Dong+2

City University of Hong Kong · School of Computer Science, Peking University · Samsung

Reasoning Segmentation Remote Sensing Image Segmentation Remote Sensing VLMs Spatial Reasoning in Language Models

Open paper AskPDF arXiv

Multimodal Large Language Models (MLLMs) have significantly advanced instruction-driven perception through the "Embedding-as-Mask" paradigm. However, extending this capability to Remote Sensing (RS) reasoning segmentation remains exceptionally challenging under cluttered geographical contexts and extreme scale variations, often leading to Spatial Fidelity Degradation and Target Misidentification. We identify a critical yet underexplored bottleneck: the sequential alienation of the segmentation token [SEG] from its initial visual grounding within the autoregressive generation paradigm. Produced only at the end of long reasoning chains, [SEG] relies on visual evidence that has been progressively diluted by linguistic abstraction and autoregressive noise. This information decay is particularly catastrophic for RS targets with minuscule pixel footprints, where even marginal fidelity loss leads to localization failure. To address these bottlenecks, we propose GeoReason, a novel framework designed to fortify logical deduction and preserve spatial integrity. First, we introduce an Anticipatory Prior mechanism that injects the [SEG] token directly into the initial user query, shifting it from a terminal output to a primary condition. This paradigm shift transforms the token into a persistent spatial anchor, preventing information decay during long-form linguistic inference. Second, we enhance visual reasoning via a Location-aware Chain-of-Thought (CoT) that enforces coarse-to-fine localization using bounding boxes, coupled with a Latent Patch-to-Token Alignment (PTA) Loss that explicitly aligns the latent segmentation token with physical RS textures. Extensive experiments demonstrate that GeoReason consistently outperforms state-of-the-art methods on the Earthreason and RRSIS-D benchmarks. Code is available at https://github.com/Fanc0126/GeoReason.

Abstract

VLM Reasoning

Explore similar work

GeoReason: Bridging Logical Reasoning and Spatial Fidelity in Remote Sensing Segmentation

GeoSearcher: Anchor-Guided Progressive Reasoning for Remote Sensing Visual Grounding with Process Supervision

Think and Answer ME: Benchmarking and Exploring Multi-Entity Reasoning Grounding in Remote Sensing