acceptodds
Under review as a conference paper at ICLR 2027

DIA-RS:Decoupled Information Alignment for Remote Sensing Reasoning Segmentation

Abstract

Existing remote sensing image (RSI) reasoning segmentation methods follow the “Embedding-as-Mask” paradigm, directly transforming high-level semantic representations from multimodal large language models into mask prompts. However, we observe that, in complex RSI scenes with high visual load, target-related semantic responses are prone to spatial displacement. We hypothesize that this phenomenon may be related to the structural coupling inherent in the existing paradigm: a single special token simultaneously undertakes target semantic representation, visual evidence localization, and mask construction, leaving the mapping from language concepts to the visual space insufficiently and implicitly modeled. To examine this hypothesis, we conduct structural intervention experiments by modifying the token coupling scheme. The results show that explicitly modeling the mapping between language concepts and the visual space can significantly alleviate the degradation of vision-language correspondence as visual load increases, further supporting the association between the proposed structural bottleneck and semantic-visual misalignment. Inspired by these observations, we propose Decoupled Information Alignment for Remote Sensing Reasoning Segmentation (DIA-RS). The method first enhances visual semantic queryability to construct a visual evidence space that is more amenable to concept querying. It then introduces an independent token together with a Concept-to-Evidence Adapter (CEA) to explicitly establish the correspondence between linguistic concepts and task-relevant visual regions. Subsequently, Evidence-guided Fusion (EGF) injects the retrieved visual evidence into , thereby enabling explicit alignment from linguistic concepts and visual evidence to mask prediction. Extensive experiments on multiple public benchmarks demonstrate that DIA-RS achieves more stable and significant performance improvements over existing baseline methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Remote Sensing Reasoning Segmentation;Remote Sensing Images;Vision-Language Spatial Alignment, Decoupled Information Alignment

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_4CBQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DIA-RS², a decoupled information alignment framework for remote sensing image reasoning segmentation. The authors argue that existing Embedding-as-Mask paradigms rely on a single <SEG> token to simultaneously undertake semantic representation, visual evidence grounding, and mask construction, which may lead to semantic-visual misalignment under high visual load scenarios. To address this, the paper designs a VSA module and introduces an independent <CON> token, a CEA, and an EGF module, establishing an explicit alignment pathway from linguistic concepts to visual evidence and ultimately to segmentation prediction.

Critical strengths
  1. Valuable research motivation. This paper focuses on the potential information alignment bottleneck between high-level semantic understanding and low-level spatial localization in remote sensing reasoning segmentation, particularly emphasizing the localization difficulties in complex scenes and under high visual load conditions. This provides a meaningful perspective for improving the existing Embedding-as-Mask paradigm.
  2. Complete and reasonable method design. The authors enhance visual feature representation capabilities via VSA, establish the correspondence between concepts and visual evidence using an independent <CON> token coupled with explicit spatial supervision, and inject the retrieved visual evidence into the segmentation prompt via EGF.
  3. Comprehensive experimental analysis. The authors provide module ablations, spatial attention analyses, evaluations across different visual load levels, cross-category generalization, and few-shot adaptation experiments, as well as a discussion of failure cases in the appendix.
Critical weaknesses
  1. Insufficient causal validation of the core hypothesis. This paper attributes semantic-visual misalignment under high visual load to the functional coupling within a single <SEG> token, but the current experiments are insufficient to prove that such coupling is the primary cause. Specifically, the G1/G2/G3 structural intervention experiments simultaneously alter spatial supervision, feature interaction, and evidence fusion mechanisms, making it difficult to cleanly attribute the performance gains solely to functional decoupling.
  2. Lack of direct quantitative evidence for "Spatial Displacement". Although Fig. 2 illustrates changes in spatial responses across different visual load levels and the appendix analyzes the relationship between attention concentration and segmentation performance, these results primarily reflect localization quality and segmentation accuracy rather than directly quantifying the "spatial displacement" phenomenon emphasized by the authors. Furthermore, since CEA itself receives explicit spatial supervision from the ground-truth mask during training, it is not surprising that its attention concentrates more on target regions.
  3. Relatively limited actual performance contribution from the core DIA module. According to Table 2, on EarthReason, introducing VSA alone improves gIoU from 72.96% to 75.34%, whereas introducing DIA alone only improves it to 73.26%; adding DIA on top of VSA yields a further gain of only 0.73 percentage points. Similar trends suggest that the overall performance gains largely stem from the relatively conventional visual feature enhancement module (VSA) rather than the decoupled information alignment mechanism emphasized in this paper.
  4. Lack of computational overhead and efficiency analysis. Building upon the LISAt framework, this paper introduces SAM LoRA, multi-level visual feature fusion, additional cross-attention, and evidence fusion modules, yet fails to report parameter counts, FLOPs, inference latency, GPU memory footprint, or training time.
  5. Insufficient targeted evaluation on high-density, multi-target, and fragmented scenarios. While the motivation emphasizes that remote sensing targets are typically small, densely distributed, and spatially fragmented, the experimental evaluation mainly relies on overall cIoU/gIoU metrics and selected qualitative visualizations, lacking independent quantitative analyses stratified by target density, object count, or fragmentation level.
Questions for authors

Please see weaknesses. I would be willing to raise my score if the authors satisfactorily address the above concerns.

Reviewer_CYz7OpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents DIA-RS², a decoupled information alignment framework for remote sensing reasoning segmentation. The paper identifies the functional coupling between semantic representation and spatial mask generation in existing Embedding-as-Mask paradigms and argues that this limitation becomes more evident under high visual load. To address this issue, the authors introduce VSA, CEA, and EGF to explicitly align semantic concepts with spatial visual evidence before mask generation. Experiments on multiple benchmarks demonstrate consistent improvements over strong baselines, including better generalization to unseen categories. Overall, the paper provides a clear motivation and a technically reasonable solution to an important problem in remote sensing reasoning segmentation.

Critical strengths

1.The paper focuses on an important but relatively underexplored issue in multimodal remote sensing segmentation: the functional coupling between semantic representation and spatial mask generation. The analysis of the Embedding-as-Mask paradigm provides a useful perspective beyond simply introducing another architectural module.

2.The method is simple and straightforward, closely aligned with the core motivation, and does not rely on complex task-specific designs, making it highly generalizable and reusable.

3.The experiments cover multiple datasets and task settings, including reasoning segmentation, referring segmentation, unseen-category evaluation, and cross-distribution generalization. The results consistently support the effectiveness of the proposed framework over the corresponding baseline.

4.The problem motivation, architecture, ablation studies, and failure analysis are presented in a reasonably coherent manner.

Critical weaknesses

1.The proposed visual-load measure combines several manually designed factors, including target scale, fragmentation, dispersion, boundary complexity, and background edge density. Although this is reasonable for constructing controlled difficulty levels, the paper could better discuss why these factors are sufficient and how sensitive the conclusions are to the particular formulation.

2.Since the proposed framework introduces additional visual features and an explicit evidence-alignment pathway, reporting parameter count, inference cost, and memory overhead would make the practical trade-off clearer.

3.Although the ablation study from G1 to G3 provides evidence that explicit spatial alignment and functional decoupling are beneficial, some improvements may also be influenced by the additional spatial supervision and intermediate evidence representation. A more strictly controlled comparison would make the causal contribution of decoupling more convincing.

Questions for authors

1.How sensitive are the conclusions to the definition of visual load? Would similar trends be observed if the five manually designed factors were weighted differently or if individual factors were removed?

2.What is the computational overhead introduced by DIA-RS² compared with the original LISAt architecture?

3.Can the authors provide a more controlled ablation to isolate the effect of functional decoupling from additional spatial supervision? In particular, it would be interesting to compare models with similar supervision and model capacity while changing only whether semantic and spatial evidence representations are explicitly separated.

Reviewer_DsZbOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a reasoning segmentation method for remote sensing images by identifying the information coupling problem in the LISA-based reasoning segmentation paradigm and introducing a decoupled architecture to address it. The paper presents a compelling research hypothesis: existing approaches rely on a single [SEG] token to simultaneously encode semantic, spatial, and segmentation-related information, and disentangling these responsibilities could improve reasoning segmentation performance. From an architectural perspective, explicitly modeling the correspondence between linguistic concepts and visual regions, and leveraging retrieved visual evidence to enhance segmentation prompts, offers an interesting and potentially novel direction beyond simply introducing additional [SEG] tokens. However, from the perspective of mechanistic validation, the paper's claims appear somewhat overstated. First, the ablation studies suggest that, compared with the LISAt baseline, the majority of the performance improvements come from VSA rather than the functional decoupling emphasized by the authors. Second, the proposed decoupling essentially separates visual evidence retrieval from mask prompt construction, rather than fully disentangling semantic, spatial, and segmentation information. Although Section 3 presents structural intervention experiments to justify the necessity of decoupling, the G2 configuration primarily demonstrates that explicit spatial supervision can improve the correspondence between linguistic concepts and visual targets. Whether introducing an independent [CON] token is necessary remains insufficiently investigated. In particular, the comparison between G2 and G3 lacks a crucial control experiment in which a single [SEG] token performs both attention supervision and visual evidence injection. Without this comparison, it is difficult to isolate the benefits of functional decoupling from those of explicit visual evidence integration.

Critical strengths
  1. The paper raises an interesting hypothesis regarding the functional coupling of semantic understanding, visual grounding, and mask generation within a single [SEG] token, providing a meaningful perspective on the limitations of existing LISA-based reasoning segmentation methods.
  2. The proposed concept-to-evidence alignment mechanism explicitly establishes the correspondence between linguistic concepts and visual regions. Incorporating the retrieved visual evidence into segmentation prompts offers an alternative to simply increasing the number of special tokens.
Critical weaknesses
  1. The ablation studies indicate that most of the performance improvements over the LISAt baseline are attributable to VSA rather than the proposed decoupling mechanism. This weakens the experimental support for the paper's central claim.
  2. The comparison between G2 and G3 introduces both an independent [CON] token and visual evidence injection. A control experiment using a single [SEG] token for both attention supervision and visual evidence injection is needed to isolate the contribution of token-level functional decoupling.
Questions for authors
  1. I noticed that recent works, such as AttnSeg (Abdelrahman et al., 2026) and the cited SegEarth-R2 (Xin et al., 2026), have explored different ways of leveraging the [SEG] token's hidden states or attention maps for downstream segmentation, with different supervision strategies and information transfer mechanisms. In this context, could the authors further clarify the motivation for constructing visual evidence through concept-conditioned cross-attention and subsequent feature aggregation, rather than directly leveraging the [SEG] token's hidden states or supervised attention maps? This question is not intended to request additional comparative experiments, but rather to better understand the rationale behind the proposed visual evidence construction strategy.

References

Segment Where You Look: Attention as the Interface Between Language Models and Segmenters. arXiv:2610.18729, 2026.

Segearth-r2: Towards comprehensive language-guided segmentation for remote sensing images. CVPR 2026

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.