acceptodds
Under review as a conference paper at ICLR 2027

UHR-Micro: Diagnosing and Mitigating the Resolution Illusion in Earth Observation VLMs

Abstract

Vision-Language Models (VLMs) are increasingly used to analyze ultra-high-resolution (UHR) Earth observation imagery, yet they face a severe scale mismatch between broad scene context and micro-scale targets. We refer to this phenomenon as a “resolution illusion”: higher input resolution provides access to more visual detail, but does not necessarily translate into reliable perception of task-relevant micro-evidence. To benchmark this challenge, we introduce UHR-Micro, a benchmark comprising 11,072 instructions grounded in 1,212 UHR images, designed to evaluate VLMs on micro-scale evidence in native-resolution Earth observation imagery. UHR-Micro spans diverse target scales, task families, and visual conditions, with each sample having an objectively verifiable target. Experiments with representative high-resolution VLMs show substantial failures in localizing and interpreting task-relevant evidence, despite access to high-resolution inputs. Further analysis shows that model scaling alone is insufficient, while localized evidence substantially improves performance, pointing to evidence access as a major bottleneck. Motivated by this finding, we propose Micro-evidence Active Perception (MAP), which constructs a spatially grounded evidence state from localized observations and supplements it when the current evidence is insufficient. Across two backbone VLMs, MAP improves UHR-Micro performance by 8.65 percentage points on average. UHR-Micro and MAP provide a framework for diagnosing and improving high-resolution reasoning in Earth observation VLMs.

Primary area:
datasets and benchmarks
Keywords:
Vision-language models, Earth observation, ultra-high-resolution imagery, resolution illusion, UHR-Micro

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_pFHeOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents UHR-Micro, a diagnostic benchmark designed to systematically evaluate micro-level perception in Earth observation VLMs, exposing a fundamental failure mode termed the "resolution illusion"—where higher input resolution fails to translate into effective micro-evidence grounding. The core contributions and empirical findings are:

  1. UHR-Micro Benchmark: As the central contribution, UHR-Micro provides 11,072 deterministic instructions grounded in 1,212 native 4K+ remote sensing images across 16 tasks. By constraining target objects to occupy under 0.01% of the canvas on average and enforcing verifiable outputs, the dataset strictly isolates spatial grounding bottlenecks from language priors and coarse scene-context shortcuts.
  2. Diagnostic Evaluation & Empirical Findings: Extensive testing across open/closed general VLMs and geospatial models reveals pervasive failures in visual evidence access. The benchmarks show that simply increasing input resolution, scaling model parameters, or adopting domain-specific tuning fails to resolve the resolution illusion.
  3. Reference Baseline (MAP): To validate that evidence access—rather than underlying recognition capacity—is the primary bottleneck, the authors introduce a simple, training-free reference baseline (Micro-evidence Active Perception, or MAP). By dynamically maintaining a spatially registered evidence state, MAP yields a consistent +8.65% average gain across backbones, confirming the benchmark's diagnostic insights on UHR reasoning.
Critical strengths
  1. A Sharp and Grounded Problem Perspective: The paper brings much-needed clarity to Earth observation VLMs by questioning a common assumption: that higher input resolution automatically yields better fine-grained perception. Pinpointing the gap between nominal resolution and actual micro-evidence grounding provides a solid, realistic perspective on the field's current bottleneck.
  2. Disciplined Benchmark Construction: UHR-Micro isolates micro-level reasoning effectively. By restricting targets to <0.01% of the canvas and enforcing deterministic outputs, the benchmark strips away language priors and dataset shortcuts, serving as a strict diagnostic stress test for spatial grounding.
  3. Sound and Practical Active Perception Paradigm: MAP offers a logical alternative to costly sliding-window search. Maintaining a spatially registered evidence state and querying additional crops only when necessary treats UHR perception as targeted evidence acquisition, yielding a consistent +8.65% boost across backbones under modest query budgets.
Critical weaknesses

While the paper is technically solid and well-executed, I would encourage the authors to clarify two minor points during the rebuttal:

  1. Part-Level Segmentation Performance and Pipeline Bottlenecks: As shown in Table 3 and Table 17, performance on Component Segmentation (CS) remains noticeably low across all evaluated models, with MAP yielding smaller relative gains compared to object-level tasks. It would be helpful if the authors could clarify whether this limitation stems from VLM feature extraction bottlenecks at the part level, or if the current Box-to-SAM2 prompt conversion protocol incurs spatial error amplification on minuscule targets.
  2. Prompt Sensitivity of the Baseline MAP: The MAP baseline uses prompt-guided queries to direct active crop selection. It would be helpful if the authors could briefly comment on the robustness of this mechanism to prompt variation—specifically, whether minor changes in prompt template phrasing impact the consistency of crop location proposals across different VLM backbones.
Questions for authors
  1. Could the authors clarify whether the lower performance on Component Segmentation (CS) is primarily driven by VLM localization errors (imprecise bounding boxes) or SAM2's prompt sensitivity when processing ultra-small target crops?
  2. In the MAP baseline, does the evidence registration accuracy vary significantly depending on whether the underlying VLM uses static patching (e.g., CLIP-based) versus dynamic resolution vision encoders (e.g., AnyRes/InternVL)?
Reviewer_THUXOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces UHR-Micro, a benchmark containing 11,072 instructions across 1,212 ultra-high-resolution remote-sensing images, covering small-target localization, recognition, counting, and spatial reasoning. It also presents MAP, a training-free inference pipeline that selects local crops, records spatially registered observations, and requests additional views when needed. Experiments report an average improvement of 8.65 points over direct inference across two VLM backbones.

Critical strengths

1.UHR-Micro addresses a meaningful evaluation problem: locating and interpreting small targets within large remote-sensing scenes. Its diverse tasks and explicit output formats support detailed analysis of different failure modes.

2.MAP improves both evaluated backbones, and its grounding gains persist under strict IoU-only evaluation. This supports the usefulness of the overall pipeline beyond the proposed continuous localization metric.

Critical weaknesses

1.Limited methodological novelty. Related capabilities have already been developed in both general-purpose and remote-sensing models: region-focused reasoning in VisCoT [1], object-centric grounded reasoning in VolCano/VoCoT [2], adaptive visual inspection in DeepEyes [3], planning and spatially grounded evidence synthesis in RSThinker [4], and on-demand zooming in GeoEyes [5]. These approaches differ in their training and tool-use mechanisms, but collectively establish substantial precedents for MAP’s ingredients. A clearer technical comparison and controlled ablations are needed to demonstrate what its source-registered evidence state and completion procedure add beyond these existing capabilities.

2.The comparisons do not adequately isolate MAP’s specific contributions. MAP uses approximately six model calls per question, compared with one for direct inference. Although sliding-window and zoom-based methods are included, an equal-budget comparison using the same backbone and aggregation procedure is missing. Consequently, the gains attributable to evidence organization and adaptive completion remain unclear.

3.The interpretation of “resolution illusion” is underdetermined. Native input means no external resizing, but the actual processed resolutions, visual-token budgets, and tile limits are not sufficiently documented. Cropping simultaneously changes detail preservation, search scope, and background interference, so the experiments do not clearly distinguish these potential bottlenecks.

References

[1] Hao Shao et al. Visual CoT: Advancing Multi-Modal Language Models with a Comprehensive Dataset and Benchmark for Chain-of-Thought Reasoning. NeurIPS, 2024. arXiv:2403.16999.

[2] Zejun Li et al. VoCoT: Unleashing Visually Grounded Multi-Step Reasoning in Large Multi-Modal Models. NAACL, 2025. arXiv:2405.16919.

[3] Ziwei Zheng et al. DeepEyes: Incentivizing “Thinking with Images” via Reinforcement Learning. ICLR, 2026. arXiv:2505.14362.

[4] Jiaqi Liu et al. Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models. ICLR, 2026. arXiv:2509.22221.

[5] Fengxiang Wang et al. GeoEyes: On-Demand Visual Focusing for Evidence-Grounded Understanding of Ultra-High-Resolution Remote Sensing Imagery. arXiv preprint arXiv:2602.14201, 2026.

Questions for authors

1.What specific mechanism distinguishes MAP from existing zoom pipelines and grounded reasoning approaches? Can the authors clarify its relationship to models such as RSThinker and consider comparisons on compatible, overlapping tasks?

2.Can MAP be compared with query-guided or fixed-grid cropping under matched observation budgets, using the same backbone, structured outputs, and aggregation procedure?

3.What processed resolutions, visual-token counts, and tile limits are used for each model? Do the main diagnostic conclusions hold when these factors are explicitly controlled?

3 comments

Dear Reviewer,

Thank you again for your thoughtful feedback. We’ve posted our response and updated the manuscript in light of your comments. We’d be happy to provide further clarification if helpful.

The evidence-use ablation in Table R2 changed my assessment. Among targets already localized correctly in Qwen’s local views, removing source-aware evidence aggregation drops final localization accuracy from 65.43% to 26.06%, with the views, registered coordinates, and call counts held fixed. MAP also outperforms self-consistency and query-guided cropping at comparable average call counts. These controls address my main concern about what MAP adds beyond cropping and extra calls. I am raising my rating to 3 (Weak Accept).

Thank you for your thoughtful feedback and for taking the time to reassess our work. Your suggestions helped us strengthen the evidence for MAP’s contribution, and we appreciate your recognition and support.

Reviewer_2XFHOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper targets a real gap in remote-sensing VLM evaluation: whether models can find and use tiny, task-critical objects inside very large scenes, rather than answering from coarse scene context. To this end, it introduces UHR-Micro, a benchmark of 11,072 instructions over 1,212 high-resolution aerial and satellite images, covering 16 tasks across grounding, fine-grained recognition, counting, and spatial reasoning, all with deterministic answers. The evaluation shows that current VLMs, including frontier APIs, perform poorly on grounding and counting tiny targets, which the authors call a “resolution illusion”; a pilot study further suggests that giving models localized crops helps far more than scaling model size. Motivated by this, the authors propose MAP, a training-free crop-and-aggregate pipeline that gathers native-resolution local observations, maps them back to the full image, and adds extra views when evidence is missing, improving two 8B backbones by about 8.7 points on average.

Critical strengths
  1. Most remote-sensing VLM benchmarks either use small images or ask scene-level questions, so they cannot tell whether a model actually found the relevant object or simply guessed from context. UHR-Micro targets objects that occupy about 0.01% of the image, and its strongest diagnostic result shows the value of this design: Gemini-3.1-Pro scores above 90 on spatial reasoning but around 4 on grounding. This is the main reason I see real value in the benchmark despite my concerns about the evaluation protocol.

  2. All 16 tasks are reduced to boxes, masks, integers, or options, scored with explicit, format-specific metrics, and the test split is image-disjoint from the development and validation splits. This avoids LLM-as-judge scoring and makes results reproducible across model families, which is essential if the benchmark is to be adopted by the community.

Critical weaknesses
  1. The central claim is that models have access to high-resolution input but fail to use it. However, Qwen3-VL caps the number of input pixels, InternVL3.5 caps the number of tiles, and commercial APIs resize images internally, so a 15-MP image is very likely never seen at native resolution. The paper does not report the effective resolution or visual-token count for any model. Its own results point the same way: InternVL3.5’s Native and Resize-1024 scores are almost identical (23.05 vs. 22.82). The near-zero grounding of GPT-5.5 and Claude (Fig. 4: about 91% background hits for GPT-5.5) may also come from forcing a [0,1000] coordinate convention these models were not trained on, rather than from a perception failure. Without these controls, the headline finding is not established, and this is the main reason for my score.

  2. The comparison with zoom-based methods is not like-for-like. DeepEyes, HiDe and GeoEyes run on Qwen2.5-VL-7B, ZoomEye on Qwen2.5-VL-3B, and many of them score exactly 0 on GD/RD/MCR, which suggests integration or format failures. The only same-backbone baseline, ZoomSearch, scores below its own direct-inference backbone (22.76 vs. 26.55). MAP also uses about six model calls per question against one for the baselines, and on region-conditioned tasks most of its gain already comes from simply cropping the given region (InternVL RD/RC: Query Crop 20.19/52.29 vs. MAP 21.20/54.44). As a result, the second contribution reads more as a reasonable engineering pipeline than as a validated method.

Questions for authors
  1. In the “Native” setting, what is the actual input each model receives? Please report the max_pixels, tile count, or visual-token count for Qwen3-VL, InternVL3.5, and the commercial APIs. If these models internally downsample 15-MP images, how do you separate the “resolution illusion” from plain downsampling? The near-identical Native and Resize-1024 scores of InternVL3.5 (23.05 vs. 22.82) make this question central.

  2. Could the near-zero grounding scores of GPT-5.5 and Claude come from the imposed [0,1000] coordinate convention rather than from perception failure? A simple sanity check would help: the same protocol on low-resolution images with large, salient objects. Relatedly, why do the instructions mix absolute-pixel coordinates (e.g., Fig. 17) and normalized coordinates (e.g., Fig. 15)?

  3. Can you re-run the zoom-based baselines on the same Qwen3-VL-8B backbone? Why does ZoomSearch, the only same-backbone baseline, score below its own direct-inference model (22.76 vs. 26.55)? Please also compare MAP with baselines given a matched call budget (about six calls), e.g., self-consistency or Query Crop combined with the global view. This would show how much of the gain comes from MAP’s design rather than from extra compute and region cropping.

  4. What are the text-only (blind), random/majority, and human scores per task? FGR results (28–42%) are close to four-way chance, and DrR/DsR targets are explicitly marked. These numbers are needed to judge which tasks truly require micro-evidence.

  5. If doable, please report human spot-check error rates for the VLM-generated tasks (BG, PDR) and inter-annotator agreement where applicable. It would also help to reconsider whether RS/CS, which are evaluated through box prompts to SAM2, should be framed as segmentation tasks.

3 comments

Dear Reviewer,

Thank you again for your thoughtful feedback. We’ve posted our response and updated the manuscript in light of your comments. We’d be happy to provide further clarification if helpful.

Thank you for the additional experiments. The coordinate checks address my concern about the grounding protocol, and the comparisons on the same backbone and at similar call budgets make a stronger case for MAP. I found the ablation using the same local observations particularly helpful in showing what MAP adds beyond cropping. My remaining concern is the interpretation of the resolution illusion, since the current results do not fully separate detail lost during preprocessing from difficulty using the detail that remains. Taking these results into account, I have raised my rating from 2 to 3.

Thank you for your positive feedback on our work and the additional experiments.

To examine how well models use the detail that remains after resizing, we created a reduced-resolution version of the UHR-Micro validation set for testing Qwen models and their derivatives. We used aspect-ratio-preserving downsampling as needed, with a 12.85 MP cap and both final dimensions divisible by 224. For each query, all methods received the same reduced-resolution image without further internal resizing or rounding.

Table R10 supports our interpretation of the resolution illusion. Models underuse fine detail even when they receive the full high-resolution image without internal downsampling. MAP substantially alleviates this limitation, although reducing the image size slightly narrows its advantage over the backbone.

Given the time constraints, we will test the remaining Qwen-based methods in the revision and clarify the distinction between preprocessing loss and use of retained detail in our discussion of the resolution illusion.

Table R10. Overall UHR-Micro validation scores (%). Native uses the original image with each model’s usual preprocessing.

Input settingQwen2-VL-7BQwen2.5-VL-7BQwen3-VL-8BDeepEyesMAP (Qwen3-VL-8B)
Native--27.6824.1936.07
Reduced resolution17.5124.2228.2125.0635.57

Thank you again for your valuable feedback and for recognizing the value of our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.