The Earth in One Gaze: Training-Free Active Focus for UHR Remote Sensing Understanding

Yao Zhang; Pengyu Dai; Wei Guo; Jian Liang; Jian Song; Yafei Ou; Hongruixuan Chen; Naoto Yokoya

cs.CVSep 23, 2026

The Earth in One Gaze: Training-Free Active Focus for UHR Remote Sensing Understanding

Authors: Yao Zhang, Pengyu Dai, Wei Guo, Jian Liang, Jian Song, Yafei Ou, Hongruixuan Chen, Naoto Yokoya

Organizations: Wuhan University

Abstract

Multimodal large language models (MLLMs) must balance local detail against scene context when interpreting ultra-high-resolution (UHR) remote sensing (RS) imagery within a limited visual-input budget. Existing selection-based methods either prune tokens and select patches through relevance scoring, or crop actively through repeated inspection. Neither strategy directly redistributes pixels within a continuous full-scene view: the first retains selected tokens or patches, and the second re-encodes a crop detached from its surroundings. Our pilot study finds that a frozen MLLM already produces useful question-guided spatial requests, yet crop-based inspection of the selected regions does not consistently improve its answers. We therefore formulate UHR understanding as a question of where to spend a fixed pixel budget. Based on this, we introduce GazeEarth, a simple-yet-effective training-free framework that couples question-guided region selection with full-scene foveated observation. The MLLM selects evidence cells from an indexed overview; a deterministic, topology-preserving warp resamples the original image onto a fixed-size canvas, enlarging their shared neighborhood while compressing the periphery; the same frozen model answers from this focused view, using at most two MLLM calls and no external selector or iterative search. Across three UHR remote sensing benchmarks and four frozen backbones, GazeEarth improves benchmark-averaged accuracy by 4.6 to 9.4 percentage points over direct answering and 3.4 to 4.3 over overview answering, outperforming task-trained methods. Our analyses show that existing MLLMs can guide where to look in UHR images on their own, and that what they can infer from the selected evidence depends on how that evidence is presented.

Figures & tables

Figure 1: Overview of GazeEarth. (i) Different observation strategies can lead to different answers. (ii) GazeEarth combines question-guided region selection with foveated observation in a continuous full-scene view. (iii) The same strategy benefits different frozen backbones. (iv) Mean-score comparison with other methods on three remote-sensing benchmarks.

Table 2

Figure 2: Gaze hit rates across three benchmarks (mean

\pm

standard deviation).

Figure 3: Overview of GazeEarth. A frozen MLLM selects evidence regions from a question and an indexed overview. A deterministic, topology-preserving warp enlarges the selected neighborhood within a fixed-size, full-scene canvas. The same model then answers from this focused observation.

Figure 4: Observation strategies on a spatial-relation question. The example compares how different methods inspect a circular track and adjacent tennis courts. GazeEarth enlarges both landmarks within the full scene and correctly identifies their relative position.

Method	Param.	TF	LRS-GRO (%) $\uparrow$				MME-RealWorld-RS (%) $\uparrow$
Method	Param.	TF	Global	Region	Object	Avg.	Color	Count	Position	Avg.
Commercial MLLMs
GPT-4o ( OpenAI, 2024 )	–	✗	31.10	12.37	21.86	21.83	34.18	15.17	37.07	28.92
Gemini-2.5-Flash ( Gemini Team, 2025 )	–	✗	57.95	23.39	30.76	35.52	55.86	33.28	51.63	47.03
General-purpose MLLMs
LLaVA-OV ( Li et al., 2025a )	7B	✗	62.43	36.18	39.92	44.42	26.81	26.14	27.57	26.85

Table 3: Main results on LRS-GRO and MME-RealWorld-RS. Scores (%) follow the respective benchmark evaluation protocols, with Avg. denoting the overall score for each benchmark. “–” in the result columns indicates an unreported score.

Method	Param.	TF	XLRS-Bench (%) $\uparrow$
Method	Param.	TF	Cnt	SC	OSR	OP	Plan	AR	CR	SR	Avg.
Commercial MLLMs
GPT-4o ( OpenAI, 2024 )	–	✗	32.50	47.67	35.20	37.47	36.00	72.00	58.50	31.67	40.16
Gemini-2.5-Flash ( Gemini Team, 2025 )	–	✗	38.75	51.67	30.40	39.10	42.00	78.00	65.50	23.33	41.66
General-purpose MLLMs
LLaVA-OV ( Li et al., 2025a )	7B	✗	31.25	35.67	25.20	45.42	24.00	76.00	59.50	43.33	41.62

Table 4: Performance (%) on XLRS-Bench across eight categories: Counting (Cnt), Scene Classification (SC), Object Spatial Relationship (OSR), Object Properties (OP), Planning (Plan), Anomaly Reasoning (AR), Complex Reasoning (CR), and Spatiotemporal Reasoning (SR).

Variant	Gaze	Q-guided gaze	Detail	Full scene	Continuous Binding	Performance (%) $\uparrow$
Variant	Gaze	Q-guided gaze	Detail	Full scene	Continuous Binding	XLRS- Bench	LRS- GRO	MME- RealWorld-RS	Avg.
Direct	0	✗	✗	✓	–	42.92	56.90	42.11	47.31
Overview	0	✗	✗	✓	–	45.13	59.99	49.01	51.38
Gaze + Crop	1	✓	✓	✗	✗	50.55	58.93	39.81	49.76
Think with Image	1	✓	✓	✓	✗	51.43	60.54	42.24	51.40
Linked-Global + Crop	1	✓	✓	✓	✓	52.56	60.25	43.39	52.07

Table 7: Ablation study on the three benchmarks. Gaze denotes the number of region-selection rounds; ”Think with Image” indicates that both global and cropped views are provided to the model.

Appendix figures & tables36 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Qwen3-VL 8B	LLaVA-v1.6 Mistral-7B	Intern-S1 mini	Commercial MLLMs API
Execution interface	Local	Local	Local	OpenRouter
Vision patch size	$16\times 16$	$14\times 14$	$14\times 14$	Not disclosed
Vision input unit	Variable-size image	$336\times 336$ tile	$448\times 448$ tile	PNG/JPEG data URL
Resolution handling	Dynamic resolution	AnyRes tiling	Dynamic tiling	Provider-managed
Default processor limit	16,777,216 pixels	4 local tiles	12 local tiles	Not disclosed
Additional global view	–	1	1 when tiled	Not disclosed

Appendix

Table 8: Backbone interfaces and inference settings. Local backbones follow their checkpoint-default visual processing, whereas GPT-4o uses provider-managed visual preprocessing. Scene-size caps apply before source-label headers and backbone-specific or provider-side preprocessing. Generation limits count new tokens.

Benchmark	Questions	Answer format	Reporting categories
LRS-GRO	9,734	Open-ended	Global, Region, Object
MME-RealWorld-RS	3,738	Multiple-choice	Color, Count, Position
XLRS-Bench	3,080	Multiple-choice	Cnt, SC, OSR, OP, Plan, AR, CR, SR

Appendix

Table 9: Benchmark composition and evaluation scope. Counts denote question–answer instances in the evaluation sets used in this work.

Observer	XLRS-Bench (%) $\uparrow$	LRS-GRO (%) $\uparrow$	MME-RealWorld-RS (%) $\uparrow$	Avg. (%) $\uparrow$
Random Focus	14.65	21.83	12.56	16.35
Question-only MLLM	41.50	44.56	42.40	42.82
SigLIP2	39.49	54.18	35.59	43.09
RemoteCLIP	33.44	45.82	32.33	37.20
Frozen MLLM	73.89	77.60	81.74	77.74

Appendix

Table 10: Spatial-request alignment with question-derived reference cells across benchmarks. We report

P(d=0)

on the eligible examples for each observer. Avg. is the unweighted mean of the three benchmark scores; higher is better.

Figure 6: Spatial-request alignment within an allowed grid distance

r

on three benchmarks. Here,

d

is the minimum Manhattan distance between selected cells and question-derived reference cells. Curves show

P(d\leq r)

on the eligible examples for each observer; sample counts refer to these examples.

Observer	Center hit (%) $\uparrow$	Box coverage (%) $\uparrow$	Answer accuracy (%) $\uparrow$
SigLIP2	52.33	47.56	64.43
RemoteCLIP	37.99	34.33	62.94
Self-MLLM	64.99	55.65	66.29
Oracle Gaze	–	–	70.20

Appendix

Table 11: Target localization and oracle-gaze answering on 537 LRS-GRO questions from 159 images without explicit location cues. The three observers share the same per-question request status and cell budget for localization. Higher is better for all percentage metrics.

Observer	Align. (%) $\uparrow$	Grid Dist. $\downarrow$	Acc.@ Aligned (%) $\uparrow$	Acc.@ Off-target (%) $\uparrow$	Spatial Acc. (%) $\uparrow$
Random Focus	14.65	1.951	39.13	38.99	39.01
SigLIP2	39.49	1.326	40.32	37.37	38.54
RemoteCLIP	33.44	1.435	41.90	35.89	37.90
Frozen MLLM	73.89	0.408	44.40	37.20	42.52

Appendix

Table 12: Spatial-request alignment and conditional answering accuracy on XLRS-Bench.

Observer	Align. (%) $\uparrow$	Grid Dist. $\downarrow$	Acc.@ Aligned (%) $\uparrow$	Acc.@ Off-target (%) $\uparrow$	Spatial Acc. (%) $\uparrow$
Random Focus	12.56	2.170	41.98	36.72	37.38
SigLIP2	35.59	1.548	49.27	36.22	40.87
RemoteCLIP	32.33	1.599	48.64	36.16	40.20
Frozen MLLM	81.74	0.280	49.43	38.81	47.49

Appendix

Table 13: Spatial-request alignment and conditional answering accuracy on MME-RealWorld-RS.

Observer	Align. (%) $\uparrow$	Grid Dist. $\downarrow$	Acc.@ Aligned (%) $\uparrow$	Acc.@ Off-target (%) $\uparrow$	Spatial Acc. (%) $\uparrow$
Random Focus	21.83	1.516	53.63	53.07	53.19
SigLIP2	54.18	0.825	57.26	51.39	54.57
RemoteCLIP	45.82	0.996	55.86	52.36	53.96
Frozen MLLM	77.60	0.314	57.65	50.00	55.93

Appendix

Table 14: Spatial-request alignment and conditional answering accuracy on LRS-GRO.

Figure 7: Category-wise accuracy gains of GazeEarth over Direct using the same Qwen3-VL-8B backbone on (a) LRS-GRO, (b) MME-RealWorld-RS, and (c) XLRS-Bench. Bars show absolute improvements in percentage points. Insets report the overall benchmark scores (Direct

\rightarrow

GazeEarth) and their differences.

Variant	Gaze	Q-guided gaze	Detail	Full scene	Binding	XLRS-Bench (%) $\uparrow$
Variant	Gaze	Q-guided gaze	Detail	Full scene	Binding	Cnt	SC	OSR	OP	Plan	AR	CR	SR	Avg.
Visual Evidence Composition
Direct	0	✗	✗	✓	–	32.50	41.67	31.40	43.92	34.00	70.00	63.50	46.33	42.92
Overview	0	✗	✗	✓	–	35.00	60.67	28.80	44.52	42.00	71.00	64.50	45.00	45.13
Random Focus	1	✗	✓	✓	✓implicit	35.00	61.33	30.60	44.10	39.00	78.00	65.50	43.33	45.42
Gaze + Crop	1	✓	✓	✗	✗	41.88	66.33	29.20	52.29	41.00	72.00	69.50	41.67	50.55

Appendix

Table 15: Category-level controlled analysis on XLRS-Bench.

Variant	Gaze	Q-guided gaze	Detail	Full scene	Binding	LRS-GRO (%) $\uparrow$
Variant	Gaze	Q-guided gaze	Detail	Full scene	Binding	Global	Region	Object	Avg.
Visual Evidence Composition
Direct	0	✗	✗	✓	–	76.25	42.03	54.77	56.90
Overview	0	✗	✗	✓	–	77.88	50.22	56.22	59.99
Random Focus	1	✗	✓	✓	✓implicit	77.84	48.78	54.95	58.97
Gaze + Crop	1	✓	✓	✗	✗	74.66	47.39	56.95	58.93

Appendix

Table 16: Category-level controlled analysis on LRS-GRO.

Variant	Gaze	Q-guided gaze	Detail	Full scene	Binding	MME-RealWorld-RS (%) $\uparrow$
Variant	Gaze	Q-guided gaze	Detail	Full scene	Binding	Color	Count	Position	Avg.
Visual Evidence Composition
Direct	0	✗	✗	✓	–	48.07	21.35	56.41	42.11
Overview	0	✗	✗	✓	–	55.54	18.35	72.39	49.01
Random Focus	1	✗	✓	✓	✓implicit	52.43	16.80	66.98	45.64
Gaze + Crop	1	✓	✓	✗	✗	62.47	22.10	34.45	39.81

Appendix

Table 17: Category-level controlled analysis on MME-RealWorld-RS.

Observer	Color	Count	Position	Avg.	$\Delta$ vs. Random
Random	52.43	16.80	66.98	45.64	0.00
SigLIP2	56.10	19.33	66.35	47.49	+1.85
RemoteCLIP	55.46	18.11	66.98	47.08	+1.44
GazeEarth	63.67	21.78	71.12	52.43	+6.79

Appendix

Table 18: Observer substitution on MME-RealWorld-RS with Qwen3-VL-8B as the fixed answerer. Best and second-best marks are assigned per column over the whole table; tied values receive the same mark.

Observer	Global	Region	Object	Avg.	$\Delta$ vs. Random
Random	77.84	48.78	54.95	58.97	–
SigLIP2	77.58	49.52	55.83	59.54	+0.58
RemoteCLIP	77.75	50.78	56.24	60.10	+1.13
GazeEarth	77.96	50.39	58.04	61.00	+2.03

Appendix

Table 19: Observer substitution on LRS-GRO with Qwen3-VL-8B as the fixed answerer.

Observer	Cnt	SC	OSR	OP	Plan	AR	CR	SR	Avg.	$\Delta$ vs. Random
Random	35.00	61.33	30.60	44.10	39.00	78.00	65.50	43.33	45.42	0.00
SigLIP2	36.25	62.67	31.40	45.12	39.00	73.00	65.50	41.67	46.10	+0.68
RemoteCLIP	37.50	61.67	29.80	45.12	38.00	77.00	65.50	43.33	45.94	+0.52
GazeEarth	44.38	66.67	37.80	53.07	49.00	76.00	74.50	55.00	53.51	+8.09

Appendix

Table 20: Observer substitution with Qwen3-VL-8B as the fixed answerer on XLRS-Bench. Ties are broken using the higher overall average.

Benchmark	Comparison	$\Delta$	95% CI	$p_{\text{Holm}}$
XLRS-Bench	Overview $-$ Direct	$+2.21$	$[+0.85,+3.54]$	$0.004$
	GazeEarth $-$ Direct	$+10.59$	$[+9.05,+12.13]$	$<0.001$
	GazeEarth $-$ Overview	$+8.38$	$[+6.85,+9.91]$	$<0.001$
LRS-GRO	Overview $-$ Direct	$+3.09$	$[+2.15,+4.03]$	$<0.001$
	GazeEarth $-$ Direct	$+4.10$	$[+3.08,+5.12]$	$<0.001$
	GazeEarth $-$ Overview	$+1.01$	$[-0.12,+2.14]$	$0.078$

Appendix

Table 21: Statistical significance of paired comparisons.

\Delta

indicates the accuracy difference (percentage points) between the first and second method. The 95% confidence intervals (CIs) are derived from paired image-cluster bootstrap.

p

-values are computed via paired image-cluster label-swap tests and adjusted across the nine comparisons using the Holm-Bonferroni method (

p_{\text{Holm}}

Variant	XLRS- Bench (s) $\downarrow$	LRS- GRO (s) $\downarrow$	MME- RealWorld- RS (s) $\downarrow$	Avg. (s) $\downarrow$	Speedup $\bm{\uparrow}$
Visual Evidence Composition
Direct	8.051	6.417	7.563	7.344	–
Overview	4.096	3.967	3.998	4.020	1.827 $\times$
Random Focus	4.687	5.124	5.019	4.943	1.486 $\times$
Gaze + Crop	3.945	4.689	4.661	4.432	1.657 $\times$
Think with Image	4.690	5.323	5.088	5.034	1.459 $\times$

Appendix

Table 22: Qwen3-VL-8B inference efficiency. Runtime is measured in seconds; lower is better. Speedup is relative to Direct; higher is better.

Variant	MLLM calls $\downarrow$	Selection (s) $\downarrow$	Visual construction (s) $\downarrow$	Answer (s) $\downarrow$	Total (s) $\downarrow$	Speedup vs. Direct $\bm{\uparrow}$
Visual Evidence Composition
Direct	1	–	–	7.563	7.563	–
Overview	1	–	0.147	3.851	3.998	1.892 $\times$
Random Focus	2	0.889	0.325	3.805	5.019	1.507 $\times$
Gaze + Crop	2	0.905	0.176	3.580	4.661	1.623 $\times$
Think with Image	2	0.884	0.252	3.952	5.088	1.486 $\times$

Appendix

Table 23: Qwen3-VL-8B inference efficiency on MME-RealWorld-RS. Lower is better for the number of MLLM calls and runtime components, while higher is better for speedup. Ties are broken using the lower total runtime.

Variant	MLLM calls $\downarrow$	Selection (s) $\downarrow$	Visual construction (s) $\downarrow$	Answer (s) $\downarrow$	Total (s) $\downarrow$	Speedup vs. Direct $\bm{\uparrow}$
Visual Evidence Composition
Direct	1	–	–	6.417	6.417	–
Overview	1	–	0.169	3.798	3.967	1.618 $\times$
Random Focus	2	1.118	0.212	3.794	5.124	1.252 $\times$
Gaze + Crop	2	1.141	0.060	3.488	4.689	1.369 $\times$
Think with Image	2	1.155	0.192	3.976	5.323	1.206 $\times$

Appendix

Table 24: Qwen3-VL-8B inference efficiency on LRS-GRO. Lower is better for the number of MLLM calls and runtime components, while higher is better for speedup. Ties are broken using the lower total runtime.

Variant	MLLM calls $\downarrow$	Selection (s) $\downarrow$	Visual construction (s) $\downarrow$	Answer (s) $\downarrow$	Total (s) $\downarrow$	Speedup vs. Direct $\bm{\uparrow}$
Visual Evidence Composition
Direct	1	–	–	8.051	8.051	–
Overview	1	–	0.331	3.765	4.096	1.966 $\times$
Random Focus	2	0.459	0.481	3.747	4.687	1.718 $\times$
Gaze + Crop	2	0.465	0.193	3.287	3.945	2.041 $\times$
Unlinked Global + Crop	2	0.474	0.422	3.794	4.690	1.717 $\times$

Appendix

Table 25: Qwen3-VL-8B inference efficiency on XLRS-Bench. Lower is better for the number of MLLM calls and runtime components, while higher is better for speedup. Ties are broken using the lower total runtime.

Figure 8: Hyperparameter sensitivity of GazeEarth with Qwen3-VL-8B: (a) selection-grid size, (b) maximum number of selected cells, (c) focus enlargement factor, and (d) selector-overview long-side cap in pixels. Solid curves show individual benchmark scores; the dashed curve shows their unweighted mean. Shaded bands mark the default settings.

Setting	Color	Count	Position	Avg.
Grid-size Sensitivity
Grid $3{\times}3$	61.75	20.88	67.78	50.37
Grid $4{\times}4$	63.67	21.78	71.12	52.43
Grid $5{\times}5$	62.15	21.78	69.29	51.31
Region-budget Sensitivity
$K=1$	63.75	22.02	70.56	52.35

Appendix

Table 26: Category-level parameter sensitivity on MME-RealWorld-RS. Bold setting names mark the default configuration. Best and second-best marks are assigned within each sweep and column; tied values receive the same mark.

Setting	Global	Region	Object	Avg.
Grid-size Sensitivity
Grid $3{\times}3$	77.23	50.44	56.99	60.28
Grid $4{\times}4$	77.96	50.39	58.04	61.00
Grid $5{\times}5$	77.45	49.78	56.85	60.11
Region-budget Sensitivity
$K=1$	77.84	50.74	56.69	60.35

Appendix

Table 27: Category-level parameter sensitivity on LRS-GRO.

Setting	Cnt	SC	OSR	OP	Plan	AR	CR	SR	Avg.
Grid-size Sensitivity
Grid $3{\times}3$	43.13	67.67	34.00	50.42	45.00	77.00	70.50	46.67	50.97
Grid $4{\times}4$	44.38	66.67	37.80	53.07	49.00	76.00	74.50	55.00	53.51
Grid $5{\times}5$	41.25	65.33	34.20	49.88	47.00	78.00	70.50	46.67	50.49
Region-budget Sensitivity
$K=1$	40.63	67.00	32.80	50.06	45.00	79.00	71.00	48.33	50.52

Appendix

Table 28: Category-level parameter sensitivity on XLRS-Bench.

Figure 9: Spatial responses and region selections under observer substitution (Part 1). Each example compares Random Focus, SigLIP2, RemoteCLIP, and GazeEarth for the same image and question. Spatial-response maps and cell-level summaries are shown alongside the recorded selections and final answers. Random Focus selects a cell without a learned response.

Figure 10: Spatial responses and region selections under observer substitution (Part 2). Each example compares Random Focus, SigLIP2, RemoteCLIP, and GazeEarth for the same image and question. Spatial-response maps and cell-level summaries are shown alongside the recorded selections and final answers. Random Focus selects a cell without a learned response.

Figure 11: Spatial responses and region selections under observer substitution (Part 3). Each example compares Random Focus, SigLIP2, RemoteCLIP, and GazeEarth for the same image and question. Spatial-response maps and cell-level summaries are shown alongside the recorded selections and final answers. Random Focus selects a cell without a learned response.

Figure 12: Spatial responses and region selections under observer substitution (Part 4). Each example compares Random Focus, SigLIP2, RemoteCLIP, and GazeEarth for the same image and question. Spatial-response maps and cell-level summaries are shown alongside the recorded selections and final answers. Random Focus selects a cell without a learned response.

Figure 13: Spatial responses and region selections under observer substitution (Part 5). Each example compares Random Focus, SigLIP2, RemoteCLIP, and GazeEarth for the same image and question. Spatial-response maps and cell-level summaries are shown alongside the recorded selections and final answers. Random Focus selects a cell without a learned response.

Figure 14: Spatial responses and region selections under observer substitution (Part 6). Each example compares Random Focus, SigLIP2, RemoteCLIP, and GazeEarth for the same image and question. Spatial-response maps and cell-level summaries are shown alongside the recorded selections and final answers. Random Focus selects a cell without a learned response.

Figure 15: Qualitative comparison of observation and answering across methods (Part 1). Each example presents the input image and question, the visual observations used by the compared methods, and their accompanying explanations and final answers. GazeEarth enlarges the question-selected neighbourhood within a continuous full-scene view, retaining the surrounding spatial context for answering.

Figure 16: Qualitative comparison of observation and answering across methods (Part 2). Each example presents the input image and question, the visual observations used by the compared methods, and their accompanying explanations and final answers. GazeEarth enlarges the question-selected neighbourhood within a continuous full-scene view, retaining the surrounding spatial context for answering.

Figure 17: Qualitative comparison of observation and answering across methods (Part 3). Each example presents the input image and question, the visual observations used by the compared methods, and their accompanying explanations and final answers. GazeEarth enlarges the question-selected neighbourhood within a continuous full-scene view, retaining the surrounding spatial context for answering.

Figure 18: Qualitative comparison of observation and answering across methods (Part 4). Each example presents the input image and question, the visual observations used by the compared methods, and their accompanying explanations and final answers. GazeEarth enlarges the question-selected neighbourhood within a continuous full-scene view, retaining the surrounding spatial context for answering.

Figure 19: Qualitative comparison of observation and answering across methods (Part 5). Each example presents the input image and question, the visual observations used by the compared methods, and their accompanying explanations and final answers. GazeEarth enlarges the question-selected neighbourhood within a continuous full-scene view, retaining the surrounding spatial context for answering.

Figure 20: Qualitative comparison of observation and answering across methods (Part 6). Each example presents the input image and question, the visual observations used by the compared methods, and their accompanying explanations and final answers. GazeEarth enlarges the question-selected neighbourhood within a continuous full-scene view, retaining the surrounding spatial context for answering.

Remote Sensing Image Understanding

Remote sensing image understanding uses AI to interpret satellite and aerial imagery for objects, land cover, change, hazards, and spatial relationships.

See all 207 papers

WeaveEarth: Structured Evidence Construction and Reasoning for Training-Free UHR Remote Sensing Understanding
Jul 11, 2026Xianzhi Ma, Shujun Wang, Xiaohan Li +3Remote Sensing Image Understanding Vision-Language Models
School of Frontier Sciences, Nanjing University, Suzhou, China · Computer Network Information Center, Chinese Academy of Sciences, Beijing, China
Beyond Zooming: Learning Multi-Tool Visual Reasoning for Ultra-High-Resolution Remote Sensing
Jul 28, 2026Fengxiang Wang, Jiangnan Huang, Mingshuo Chen +8RL for Tool Use Remote Sensing VQA
National University of Defense Technology · Wuhan University · Tsinghua University
GeoVista: Visually Grounded Active Perception for Vision-Language Understanding of Ultra-High-Resolution Remote Sensing Images
May 14, 2026Jiashun Zhu, Ronghao Fu, Jiasen Hu +3Remote Sensing Image Understanding Remote Sensing VQA
College of Computer Science and Technology, Jilin University, Changchun 130021, China · Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Changchun 130021, China · Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100000, China +1

Jul 11, 2026J/K move · Enter open · S save

WeaveEarth: Structured Evidence Construction and Reasoning for Training-Free UHR Remote Sensing Understanding

Xianzhi Ma, Shujun Wang, Xiaohan Li, Hao Liu+2

School of Frontier Sciences, Nanjing University, Suzhou, China · Computer Network Information Center, Chinese Academy of Sciences, Beijing, China

Remote Sensing Image Understanding Vision-Language Models VLM Reasoning

Open paper AskPDF arXiv

Ultra-High-Resolution (UHR) remote sensing image understanding requires Vision-Language Models (VLMs) to capture both the global scene layout and sparse yet task-critical local details under limited computational budgets. Existing methods mainly follow two paradigms. One is passive perception, which relies on resolution expansion or token compression and may therefore discard fine-grained details. The other is active perception, which depends on multi-round zooming and search, but suffers from high latency, contextual fragmentation, and error accumulation. We argue that a more effective path toward UHR understanding lies not in accessing more, but in organizing better. To this end, we propose WeaveEarth, a training-free framework that reformulates UHR understanding as a problem of structured evidence construction and reasoning under global context constraints. Specifically, WeaveEarth first employs Global-Aware Evidence Construction to select a compact, low-redundancy, and spatially complementary Minimal Support Evidence Set. It then introduces Structured Evidence Reasoning, which weaves local evidence, spatial metadata, and relative topology into a unified reasoning interface, thereby enhancing the VLM's ability to perform global-local joint reasoning. Extensive experiments show that WeaveEarth consistently outperforms strong baselines and existing UHR methods across multiple UHR remote sensing benchmarks and multiple frozen VLM backbones. Code is available at https://github.com/XianZhi-Ma/WeaveEarth.

Abstract

Remote Sensing Image Understanding

Explore similar work

WeaveEarth: Structured Evidence Construction and Reasoning for Training-Free UHR Remote Sensing Understanding

Beyond Zooming: Learning Multi-Tool Visual Reasoning for Ultra-High-Resolution Remote Sensing

GeoVista: Visually Grounded Active Perception for Vision-Language Understanding of Ultra-High-Resolution Remote Sensing Images