RS-OPSD: Reliable Privileged On-Policy-Self-Distillation for Ultra-High-Resolution Remote Sensing VQA

Chengjie Jiang; Yunqi Zhou; Jiafeng Yan; Sihang Zhao; Chun Yuan; Jing Li

cs.CVSep 29, 2026

RS-OPSD: Reliable Privileged On-Policy-Self-Distillation for Ultra-High-Resolution Remote Sensing VQA

Authors: Chengjie Jiang, Yunqi Zhou, Jiafeng Yan, Sihang Zhao, Chun Yuan, Jing Li

Organizations: Tsinghua University · Zhejiang University · Central University of Finance and Economics · East China Normal University · Key Laboratory of Geographic Information Science

Abstract

Ultra-high-resolution (UHR) remote sensing visual question answering (VQA) requires models to resolve small visual evidence within extremely large images. Existing approaches typically rely on token pruning, visual search, or tool-augmented reasoning at inference time. We instead investigate whether the benefit of zoom-in visual privilege can be internalized into the model. We introduce RS-OPSD, a reliable privileged on-policy self-distillation (OPSD) framework for UHR remote sensing VQA. To provide high-quality privileged information with explicit question-relevant evidence, we construct GeoEvidence-6K, containing 6,750 VQA samples across seven task categories with evidence-region annotations, and develop Human Feedback-Guided Skill Refinement (HF-SR) for scalable annotation. To address context loss from tight crops and conflicting signals from imperfect teachers, RS-OPSD introduces Context-Preserving Visual Privilege (CPVP) and Correctness-Aligned Distillation (CAD). Without any additional visual search or tool calls at inference time, RS-OPSD achieves state-of-the-art (SOTA) performance on XLRS-Bench, MME-RealWorld-RS, and LRS-VQA, outperforming previous SOTA models of comparable scale by an average of 4.0 percentage points. Moreover, our 2B variant, RS-OPD-Lite, surpasses most 8B-scale models while achieving the fastest measured inference speed. Our Code, GeoEvidence-6K, and the model weights for RS-OPSD and RS-OPD-Lite are publicly available.

Figures & tables

Figure 1: Average scores on Ultra-High-Resolution remote sensing VQA benchmarks, including XLRS-Bench, MME-RealWorld-RS, and LRS-VQA.

Figure 2: Comparison of representative paradigms for UHR remote sensing VQA.

Source	Nums.	Image Size	Target Size	GSD	Country
MiniFrance ( Castillo-Navarro et al., 2022 )	110	$10K$	$509\!\times\!517$	$0.50\,\mathrm{m}$	France
HRSCD ( Daudt et al., 2019 )	40	$10K$	$516\!\times\!517$	$0.50\,\mathrm{m}$	France
SWISSIMAGE ( Federal Office of Topography swisstopo, 2024 )	300	$10K$	$837\!\times\!817$	$0.10\,\mathrm{m}$	Switzerland
GeoNRW ( Geobasis NRW, 2025 )	300	$10K$	$997\!\times\!987$	$0.10\,\mathrm{m}$	Germany
Beeldmateriaal ( Beeldmateriaal Nederland, 2025 )	300	$12K$	$1{,}741\!\times\!1{,}696$	$0.08\,\mathrm{m}$	Netherlands

Table 1: Statistics of the image sources used in GeoEvidence-6K.

Figure 3: Overview of the HF-SR pipeline, together with representative cases of GeoEvidence-6K.

Figure 4: Overview of RS-OPSD , consisting of CPVP and CAD.

Closed-Source Vision Language Models
Method	Pub.	Backbone	Param.	XLRS	MME	LRS	Avg.	Speed
GPT-4o ( Hurst et al., 2024 )	–	–	–	32.4	28.2	27.1	29.2	-
Claude 3.7 Sonnet ( 5 )	–	–	–	40.5	45.7	28.4	38.2	-
Gemini 2.5 Pro ( Comanici et al., 2025 )	–	–	–	44.1	51.7	30.9	42.2	-
Open-Source Vision Language Models
LLaVA-OV-7B ( Li et al., 2024 )	TMLR’25	LLaVA-OV	7B	43.4	53.3	27.3	41.3	2.38

Table 2: Quantitative comparison results. Green , yellow , and blue denote the first-, second-, and third-best results. Speed is the average inference time across the three benchmarks (s/sample).

Core Design			Label	Benchmarks
OPSD	CPVP	CAD		XLRS	MME-RW-RS	LRS-VQA	Avg.
Qwen3-VL-8B-Instruct			Base Model	50.5	41.9	30.1	40.8
✗	✗	✗	LoRA Finetuning on GeoEvidence-6K	49.9	55.2	32.1	45.7
✓	✗	✗	Naïve OPSD in Subsec. 5.1	50.5	53.8	31.3	45.2
✓	✓	✗	OPSD + CPVP	51.3	59.0	31.2	47.2
✓	✗	✓	OPSD + CAD	51.2	58.3	31.5	47.0

Table 3: Ablation study of the core designs in RS-OPSD.

KL Coefficient		Benchmarks				Clipping Threshold		Benchmarks
$\bm{c_{\mathrm{grad}}}$	$\bm{\lambda_{\mathrm{KL}}}$	XLRS	MME-RW-RS	LRS-VQA	Avg.	$\bm{c_{\mathrm{grad}}}$	$\bm{\lambda_{\mathrm{KL}}}$	XLRS	MME-RW-RS	LRS-VQA	Avg.
1	0	51.3	59.0	31.2	47.2	5	$1\times 10^{-3}$	53.1	61.5	33.3	49.3
1	$1\times 10^{-2}$	51.2	62.0	32.8	48.7	20	$1\times 10^{-3}$	51.0	60.0	34.0	48.3
1	$1\times 10^{-3}$	53.2	60.8	33.1	49.0	100	$1\times 10^{-3}$	51.5	59.9	33.4	48.3

Table 4: Hyperparameter studies of KL regularization and gradient clipping.

Figure 5: Training dynamics of RS-OPSD across three epochs.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

\mathcal{L}_{\mathrm{CAD}}=-\frac{\sum_{i,t}m^{\mathrm{ans}}_{i,t}\,\operatorname{sg}[A_{i,t}]\log p^{S}_{i,t}}{\sum_{i,t}m^{\mathrm{ans}}_{i,t}}

Appendix

Algorithm 1 Training procedure of RS-OPSD

Figure 6: Dataset statistics of GeoEvidence-6K. Left: word cloud of the question texts. Middle: token-length distribution of the questions, with the mean, median, and 95th percentile marked by dashed lines. Right: distribution of task categories.

Figure 7: Prompt templates for the student and privileged teacher in RS-OPSD .

Closed-Source Vision Language Models
Method	Perception								Reasoning					Avg.
Sub-tasks	OC	RC	OLUC	RLUC	OCC	OCL	OMS	OSR	AD	ECR	RP	RCCD	CCR
GPT-4o	25.0	32.0	15.0	66.0	9.5	11.3	11.7	24.6	73.0	73.0	35.0	20.0	25.0	32.4
Claude 3.7 Sonnet	27.6	22.7	17.4	68.4	30.5	29.9	63.6	27.6	64.8	78.4	34.5	27.8	32.6	40.5
Gemini 2.5 Pro	50.0	42.0	12.0	65.5	37.5	38.1	60.0	30.2	68.0	72.0	36.0	26.7	35.0	44.1
Open-Source Vision Language Models

Appendix

Table 5: Detailed quantitative results on XLRS-Bench across perception and reasoning sub-tasks. Green , yellow , and blue denote the first-, second-, and third-best results, respectively.

Closed-Source Vision Language Models
Method	MME-RealWorld-RS				LRS-VQA				XLRS-Bench	Total
Sub-set	Position	Color	Count	Avg.	FAIR	Bridge	STAR	Avg.	Avg.	Accuracy	Speed
GPT-4o	36.4	32.4	15.9	28.2	22.2	31.8	27.4	27.1	32.4	29.2	–
Claude 3.7 Sonnet	56.6	52.1	28.5	45.7	25.7	29.5	30.1	28.4	40.5	38.2	–
Gemini 2.5 Pro	60.6	60.3	34.3	51.7	28.9	31.2	32.7	30.9	44.1	42.2	–
Open-Source Vision Language Models

Appendix

Table 6: Detailed quantitative results on MME-RealWorld-RS, LRS-VQA, and XLRS-Bench. Average inference speed is measured in s/sample.

Explore similar work

A Unified Framework for Efficient Remote Sensing Visual Question Answering: Adapting Dual, Hybrid, and Encoder-Decoder Architectures
Jun 17, 2026Timothy Agboada, Shikha Chandel, Yadav Raj Ghimire +1Remote Sensing Recent Vision-Language Models
Computational Data Science and Engineering North Carolina A&T State University Greensboro - NC, USA · College of Science and Technology North Carolina A&T State University Greensboro - NC, USA
One Adapter, Every Resolution: Gated Low-Rank Adaptation for Remote Sensing VLMs
May 8, 2026Song Zhang, Yanlong Chen, Yining Chen +2Remote Sensing Recent Vision-Language Models
Nanjing University · ETH Zurich · Nanjing University of Posts and Telecommunications
Beyond Zooming: Learning Multi-Tool Visual Reasoning for Ultra-High-Resolution Remote Sensing
Jul 28, 2026Fengxiang Wang, Jiangnan Huang, Mingshuo Chen +8Trustworthy Earth Observation Decisions Visual Reasoning
National University of Defense Technology · Wuhan University · Tsinghua University

Jun 17, 2026, cs.CVJ/K move · Enter open · S save

A Unified Framework for Efficient Remote Sensing Visual Question Answering: Adapting Dual, Hybrid, and Encoder-Decoder Architectures

Timothy Agboada, Shikha Chandel, Yadav Raj Ghimire, Leila Hashemi-Beni

Computational Data Science and Engineering North Carolina A&T State University Greensboro - NC, USA · College of Science and Technology North Carolina A&T State University Greensboro - NC, USA

Remote Sensing Recent Vision-Language Models Remote Unified Framework

Open paper AskPDF arXiv

Visual Question Answering (VQA) in the Remote Sensing (RS) domain presents unique challenges due to the high resolution, multi scale object distribution, and semantic complexity of aerial imagery. While general domain Foundation Models have achieved remarkable success, their direct application to RSVQA is hindered by massive domain shifts and the computationally prohibitive nature of full fine tuning. This study presents a comparative analysis of RS Adapter, a Parameter Efficient Fine Tuning (PEFT) strategy, applied across three distinct Vision Language Model (VLM) architectures: the Dual Encoder CLIP, the Encoder Decoder BLIP, and the Hybrid FLAVA. We introduce a unified architectural surgery pipeline that injects lightweight bottleneck adapters into the attention and MLP layers of frozen backbones, enabling rapid adaptation with less than 5 percent of trainable parameters. Experimental results on the high resolution RSVQA x dataset demonstrate that while all adapted models achieve convergence, the Hybrid FLAVA architecture offers a superior balance of multimodal reasoning and retrieval capabilities compared to its unimodal counterparts. Our findings establish a new baseline for resource efficient VQA in disaster assessment and urban monitoring.

Abstract

Explore similar work

A Unified Framework for Efficient Remote Sensing Visual Question Answering: Adapting Dual, Hybrid, and Encoder-Decoder Architectures

One Adapter, Every Resolution: Gated Low-Rank Adaptation for Remote Sensing VLMs

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