acceptodds
Under review as a conference paper at ICLR 2027

GEOEPS: TRAINING-FREE EVIDENCE PRESENTATION AND SELECTION FOR ULTRA-HIGH-RESOLUTION REMOTE SENSING UNDERSTANDING

Abstract

Ultra-high-resolution (UHR) remote-sensing (RS) imagery poses a fundamental challenge for vision-language models (VLMs): task-relevant evidence may occupy only a tiny fraction of a large scene, yet making that evidence more accessible does not ensure its effective use in the final answer. We identify this discrepancy as an evidence-to-decision gap, which is often observed in visual question answering (VQA) tasks for UHR-RS imagery. Through controlled comparisons, we find that how a referenced region is presented matters beyond simply locating it. To address this gap, we introduce GeoEPS, a training-free framework that coordinates evidence presentation and selection with frozen VLMs. When the input provides an explicit spatial cue, Evidence-aware Presentation (EAP) uses it as an evidence prior and combines local context, textual alignment, unfilled marking, and query-conditioned presentation. For inputs without an explicit spatial cue, Competitive Evidence Selection (CES) searches local evidence candidates and compares them under competing answer hypotheses. On XLRS-Bench and RSHR-Bench, our GeoEPS substantially improves Qwen3-VL-8B's average task accuracy from 48.8% to 61.0% and from 41.3% to 53.7%, respectively. Extensive experiments further demonstrate that these gains can extend to other frozen general-purpose and RS-specific VLMs without parameter updates or UHR-specific retraining.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
remote sensing, vision-language models, ultra-high-resolution imagery, visual question answering, training-free inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_F5yyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoEPS, a training-free framework for ultra-high-resolution remote-sensing VQA tasks. GeoEPS claims that making relevant evidence accessible (by cropping) is not enough, and therefore proposes EAP and CES to priors building and evidences searching, respectively. On XLRS-Bench and RSHR-Bench, it lifts Qwen3-VL-8B from 48.8% to 61.0% and from 41.3% to 53.7%, with consistent gains across other frozen VLMs.

Critical strengths
  1. The paper's motivation and identified problem are clear, and are well supported by ablations.
  2. The proposed method is training-free, architecture-agnostic, and easy to integrate.
  3. The performance gains are consistent across multiple models and benchmarks.
Critical weaknesses
  1. The proposed method incur additional inference costs in model calls, and no runtime or memory costs are reported. If GeoEPS and other methods are evaluated under a fixed model calls, will the gains hold still?
  2. CES only adds little for open-source MLLMs in Table 3, while effectively boosts performance for remote sensing MLLMs. Why?
  3. In Sec. B.3, it seems that the query-conditioned selection relies on three regex rules fitted to the benchmark question templates. Can it generalize?
Questions for authors

Please see weaknesses

Reviewer_eeqHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GeoEPS addresses a problem in ultra-high-resolution (UHR) remote sensing visual question answering (VQA): even when relevant image regions are provided to a vision-language model (VLM), the model may still use the evidence incorrectly. The authors describe this issue as an evidence-to-decision gap, where evidence is accessible but does not effectively influence the final answer. The proposed method does not require updating model parameters. When spatial hints are available, EAP improves image cropping, target reference, and view presentation. When spatial hints are unavailable, CES searches local regions and compares evidence supporting different answers. Using Qwen3-VL-8B, the method improves task-level macro-average accuracy by 12.2 and 12.4 percentage points on XLRS and RSHR, respectively.

Critical strengths
  1. The paper highlights that providing relevant image regions does not guarantee that a model will use the evidence correctly. It also introduces specific controls to address issues related to context, target identification, and visual distractions.
  2. The paper proposes EAP, which combines context-aware cropping, coordinate-text alignment, outline markers outside bounding boxes, and question-based selection of marked, clean, or paired views. For the 1,660 numeric-box questions in XLRS evaluated with Qwen3-VL-8B, the full EAP achieves a Macro₃ score of 90.2%, compared with 74.1% for the cropping baseline with correct coordinates. In a rendering control experiment using the same crops and text, the outline-only setting achieves 79.3%, while the setting without markers achieves 57.2%.
  3. The paper proposes CES, which first ranks evidence by sufficiency and then searches for evidence supporting alternative answers. It compares how well different evidence distinguishes between competing hypotheses. This is an interesting design, although its performance gain is relatively small.
Critical weaknesses
  1. On the overall XLRS task-level macro-average accuracy with Qwen3-VL-8B, EAP provides a gain of 10.8 percentage points, while CES adds only 1.4 percentage points. The additional improvement from CES is relatively small. Moreover, the paper does not report how many incorrect answers are corrected or how many correct answers become incorrect during competitive selection. Apart from the top-SV incumbent (C3), no alternative selection strategies are compared under the same candidate conditions.
  2. The paper does not report actual computational costs and only provides the number of model calls. The computational requirements of complex spatial reasoning are not discussed in sufficient detail.
  3. A large share of the benchmark questions already supply explicit bounding boxes, and most of the overall gain comes from exploiting these supplied cues through EAP. The reported evidence-to-decision gap therefore largely reflects the native model's difficulty in mapping full-resolution coordinates onto a downsampled image, rather than a deeper failure in evidence use. It is also unclear whether the compared UHR methods are given the same box-based cropping, which makes the comparison potentially unfair.
  4. The query-conditioned presentation in EAP relies on hand-crafted text-pattern rules that directly correspond to the query families of the evaluated benchmark. These rules appear to be designed and evaluated on the same test questions without a separate validation split, which raises concerns about test-set tuning and limits generalization to other question types.
Questions for authors
  1. What is the actual computational cost of the performance improvements? Please report the average number of model calls and inference costs for Native, EAP, C3, and C4 on the same hardware, with specific numerical results.
  2. What errors does DV actually correct? Please provide detailed paired statistics on incorrect-to-correct and correct-to-incorrect answer changes, together with confidence intervals. Please also compare DV with alternative selection rules using the same candidates.
Reviewer_f6WmOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoEPS, a training-free framework for ultra-high-resolution remote-sensing VQA. It introduces Evidence-aware Presentation (EAP) to improve the use of provided spatial references and Competitive Evidence Selection (CES) to search and compare evidence when such references are absent. Experiments on two benchmarks and twelve frozen VLMs report consistent overall accuracy improvements, with Qwen3-VL-8B achieving Macro gains of 12.2 and 12.4 percentage points.

Critical strengths
  • The paper highlights an important distinction between locating relevant visual evidence and effectively using it for prediction. The proposed training-free framework can be applied to frozen VLMs without additional training.
  • Experiments across twelve models and two benchmarks demonstrate consistent improvements in aggregate accuracy, including gains for remote-sensing-specific models.
  • The paper provides progressive ablations and controlled comparisons of cropping, reference alignment, visual marking, and evidence selection, helping examine the contributions of different components.
Critical weaknesses
  • GeoEPS largely combines existing techniques such as contextual cropping, visual prompting, heuristic query routing, and VLM-based verification. The paper does not sufficiently clarify its methodological novelty over existing training-free visual search and evidence-presentation approaches.
  • Each CES candidate includes both the same global overview and a local crop, making it unclear whether self-verification scores reflect the local evidence or information already available globally. Ablations removing the overview or replacing local crops with irrelevant regions are needed to validate the evidence-ranking mechanism.
  • CES selects a single local evidence candidate instead of combining complementary observations from different image regions. This design is inherently limited for counting, spatial relationships, and scene-level reasoning tasks that require information distributed across multiple locations.
  • Full CES may require up to multiple VLM calls per question, yet the paper provides no measured latency, token consumption, or compute-matched comparison. Given the small incremental gain from competitive selection, its practical cost-effectiveness remains unclear.
  • The two modules operate as mutually exclusive inference branches rather than a unified evidence-processing pipeline. The framework neither applies EAP to CES-discovered regions nor uses CES to recover from unreliable spatial cues, limiting the claimed complementarity between evidence presentation and selection.
  • Although the paper emphasizes the evidence-to-decision gap, it mainly evaluates final answer accuracy without directly assessing evidence localization, relevance, or utilization.
  • EAP relies on hand-crafted query patterns that exactly match the evaluated question templates, while CES uses fixed search heuristics. The lack of evaluation on unseen question formulations and more diverse VQA settings limits confidence in the framework's general applicability.
Questions for authors

Please see the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.