acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Remote sensing imagery spans ground sampling distances (GSDs) from centimeters to tens of meters, so both the visual evidence for a geographic concept and the questions it can support change with physical scale. Existing remote sensing vision-language models (RS-VLMs) either ignore GSD or encode it as a text token, applying one set of scale-agnostic parameters across this spectrum. We show that GSD is decodable from frozen visual features, but exploiting it requires scale-dependent adaptation. We therefore introduce ScaleEarth, which conditions low-rank adaptation on the continuous scalar . Its core module, CS-HLoRA, gates the rank dimensions of a single shared LoRA with sigmoid functions of whose thresholds are initialized at object, structure, and semantic scales; the gates route gradients by scale during training and select the active adaptation subspace at inference, with the backbone frozen. When metadata is unavailable, a heteroscedastic head, SSE-U, estimates with calibrated uncertainty and falls back to a default scale when uncertain. To align supervision with this mechanism, we build GeoScale-VQA, 1.5M QA pairs in which every sample carries a resolved GSD and newly generated questions are conditioned on the same . With an 8B backbone, ScaleEarth achieves an average score of 59.4 on XLRS-Bench ( over the strongest RS specialist). On OmniEarth-Bench, it achieves an average score of 40.71 ( over the strongest open-source baseline). The largest gains occur on scale-sensitive sub-tasks. The 4-bit variant can be deployed on a single 40 GB GPU and retains an average score of 57.4 on XLRS-Bench, with substantially lower adaptation and deployment costs.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Remote sensing, VLM, Efficient AI

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Y69COpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work considers parameter-efficient adaptation of remote-sensing vision-language models across imagery with different ground sampling distances (GSDs). The proposed ScaleEarth framework uses GSD to condition the low-rank update of a frozen backbone. Its main component, CS-HLoRA, inserts a GSD-dependent gate between the two LoRA projections, so that different rank dimensions are activated at different input scales. These dimensions are initialised in three groups, referred to as object, structure, and semantic ranks. An additional module, SSE-U, estimates GSD from visual features when metadata are unavailable. The authors also construct GeoScale-VQA, a collection of roughly 1.5M question-answer pairs with associated GSD information. The manuscript presents this combination as a means of making physical scale directly influence model adaptation rather than treating GSD as passive metadata. Results are reported on XLRS-Bench and OmniEarth-Bench. Besides comparisons between complete VLM systems, the experimental section includes controls using standard LoRA, textual GSD input, FiLM, GSD-routed MoE variants, and gated/adaptive PEFT methods

Critical strengths

The experimental section makes a serious effort to control several simple confounders. The contribution is not supported solely by comparisons between complete VLM systems. A number of Stage-2 adaptation mechanisms are compared from the same Stage-1 checkpoint, with common training data and training budgets

The GSD-shuffling experiment is particularly useful. Breaking the image–GSD correspondence while retaining the corpus, architecture, parameter count and training budget causes a substantial deterioration. This provides good evidence that the correct GSD association carries useful information rather than merely reflecting additional data or capacity

The factorial experiment separating the training corpus from the adaptation mechanism is also helpful. These controls make me believe that the empirical gains are genuine. My concern is that they establish a narrower result than the one claimed

Critical weaknesses

My main difficulty is that the methodological contribution is presented at a level of interpretation which the experiments do not establish. At its core, CS-HLoRA is a conditional parameter-efficient adaptation scheme. Relative to standard LoRA, \Delta W=alpha/r BA, the essential modification is \Delta W(s)=alpha/r B diag(h(s))A, where the external scalar s is derived from GSD. Thus, GSD controls which dimensions of the low-rank update participate for a given sample. The ranks are then initialised in three groups and labelled object, structure, and semantic. This is a conditional low-rank adaptation rule; I do not see a physical model from which this particular parameterisation follows.

There is an important distinction here. GSD has a clear physical meaning, and spatial sampling plainly affects the information available in an image. It does not follow that the low-rank parameter space of a VLM possesses a corresponding object/structure/semantic hierarchy. Nor does it follow that particular rank dimensions should represent these functions. The gate is interpretable in the limited sense that its value depends on GSD. The parameters being gated remain black-box parameters learnt end-to-end. A physically meaningful conditioning variable does not, by itself, make the conditioned computation physically interpretable

Relatedly, I could not find direct evidence that the three rank groups acquire the functions assigned to them. If object ranks genuinely specialise in fine-scale object evidence, suppressing them should selectively damage fine-grained object tasks. Likewise, removing semantic ranks should have a different and predictable effect on coarse scene-level reasoning. Such a functional dissociation would provide evidence for the proposed hierarchy. At present, the results mainly show that GSD-dependent training changes the learnt parameters and improves aggregate performance. This is not the same as demonstrating object/structure/semantic specialisation

A second difficulty concerns the status of GSD itself. The experiments show that the correct image–GSD association is useful, but do not show that the benefit is specific to the physical meaning of GSD. GSD is potentially correlated with sensor, source dataset, image statistics, scene distribution, and the distribution of questions. GeoScale-VQA is itself organised using scale information. A simpler interpretation therefore remains possible: heterogeneous remote-sensing data benefit from conditioning the adaptation procedure on an informative sample-level covariate

This distinction matters. If sensor identity, source-domain identity, image complexity, or a learnt latent condition were to provide a similar advantage when used with the same gating mechanism, the result would primarily concern conditional adaptation rather than physical-scale-driven modelling. Shuffling GSD does not distinguish these explanations. It shows that the correct association matters, but not why it matters

There is also a mismatch between the most prominent comparisons and the actual algorithmic contribution. Much of the headline evidence concerns complete ScaleEarth systems compared with other RS-VLMs. Those comparisons are useful for judging the resulting system, but differences in training data, adaptation recipe, GSD conditioning, SSE-U and other choices prevent them from identifying the merit of CS-HLoRA itself. The additional LoRAMoE, MoLE and GSD-routed MoE-LoRA comparisons are much more relevant. They show that the proposed continuous rank gating performs well against several alternative adaptation strategies. They still do not establish the physical interpretation placed upon that result

A simple substitution test illustrates my concern. The mathematics of CS-HLoRA remains intact if GSD is replaced by another continuous conditioning variable. Conversely, GSD could condition another adapter architecture without changing the underlying motivation. There is therefore no evident necessity linking the remote-sensing quantity to this particular adaptation rule. A more precise description of the contribution would be GSD-conditioned low-rank adaptation for remote-sensing VLMs. That is appreciably narrower than a physically grounded adaptation mechanism

The experiments support the statement that GSD is a useful conditioning variable for parameter-efficient adaptation in this setting. I do not think they presently support the stronger statement that physical scale induces the proposed hierarchy of adaptation subspaces, and this hierarchy explains the benefit of the proposed routing mechanism. The distinction between these two statements is central to my assessment

Questions for authors

What evidence shows that GSD is a privileged conditioning variable rather than simply an informative domain covariate? Under the same adapter architecture and training protocol, it would be useful to compare GSD with sensor identity, source-domain identity, image-complexity measures, or a learnt latent condition. Comparable gains from such alternatives would substantially change the interpretation of the method

Is there direct evidence that the proposed rank groups acquire object-, structure-, and semantic-level functions? A group-by-task intervention would be informative: suppress each rank group independently and examine whether the resulting failures exhibit the functional selectivity implied by the proposed interpretation

Can physical scale be varied while holding scene, sensor, semantic content and question approximately fixed? For instance, a controlled resolution degradation experiment on the same imagery would test whether the preferred gating changes systematically with actual loss of resolvable detail. This would provide considerably cleaner evidence than correlations across heterogeneous sources

How do the authors distinguish the claim that “GSD is an informative conditioning variable” from the considerably stronger claim that “GSD physically determines the appropriate adaptation subspace”? The shuffling, spoofing and parameter-matched experiments are useful evidence for the former. I do not see an experiment that identifies the latter.

7 comments
AuthorsOpenReview ↗

Response to Reviewer Y69C

Thank you for your thorough and insightful review. We have carefully addressed your questions in the revised manuscript and detail the changes below. If these updates resolve your concerns, we would be grateful if you could consider a more favorable rating.

W1 / Q4. Is CS-HLoRA only a conditional adapter with a physical label? Our new tests go beyond "GSD is an informative covariate". Sec. 3.2 (lines 323-325) states our shared premise: sampling physics motivates the gate rather than deriving it. Our claim is empirical and rests on three results:

  • With scene, sensor, question and weights fixed, the preferred gate setting follows the image's physical resolution (W3; Table 18, lines 2147-2152).
  • The three rank groups fail selectively on their own task families (W2; Table 17, lines 2085-2093).
  • The nesting that the motivation predicts is measurable. A nested adapter with a hard gate scores 39.31 on OmniEarth-Bench vs. 37.89 for disjoint GSD-bucketed experts; smoothing adds 1.40 to reach 40.71 (App. G.4, lines 2186-2209; Table 19).

We ran your substitution test in both directions. Four other variables through the same gate recover at most 37% of the gain (Table 16, lines 2029-2035). GSD through three other mechanisms reaches 32.60 (prompt), 36.28 (FiLM) and 37.89 (bucketed experts) vs. 40.71 for rank gating (Tables 2, 7). While the rule accepts any scalar, none tested delivers the gain of GSD.

W2 / Q2. Do the rank groups acquire object-, structure- and semantic-level functions? Yes. Your proposed intervention reveals a clear double dissociation across all three groups (App. G.2, lines 2071-2129; Table 17). On the full XLRS-Bench we zero one group's gate in the trained 4-bit model and compare with removing the same number of random ranks (mean of five draws) across pre-fixed task families. Zeroing the object group (ranks 0-10) drops object tasks by 4.6 points vs. 1.1 for 11 random ranks; zeroing the structure group (ranks 11-31) drops structure tasks by 4.9 vs. 1.9 for 21 random ranks; and zeroing the semantic group (ranks 32-63) drops scene tasks by 7.4 vs. 3.1 for 32 random ranks. Each group hurts its own family 2.4 to 4.2 times more than a size-matched random set, while every off-diagonal loss stays at or below its control (0.5 to 2.6 points). The effect also follows scale: removing the object group costs 8.4 points on object tasks on Toronto (0.05-0.1 m), 5.0 on HRSCD (0.5-1 m) and 0.9 on DOTA-v2.0 (2-4.5 m), where the gate has already closed it. This pattern is inconsistent with arbitrary labels on black-box parameters. App. A.9 (lines 1387-1389) now frames the earlier probe as correlational and this intervention as the causal test.

W3 / Q1 / Q3. Is GSD privileged, and is it scale or source? GSD outperforms every alternative tested, and its effect holds with the source fixed. We test both predictions of the "informative covariate" account:

Prediction 1: source or sensor identity should do as well (App. G.1, lines 2012-2068; Table 16; 300K samples, three seeds, same adapter). While log-GSD raises OmniEarth / XLRS from 31.8 / 52.4 (standard LoRA) to 38.0 / 55.8 (100% of the gain), feeding learned latent scalars (33.4 / 53.2), image complexity (33.9 / 53.5), sensor identity (33.6 / 53.3) or source identity (34.1 / 53.6) to the same gate recovers only 26% to 37% of the GSD gain. Source identity matches GSD only on held-out GeoScale-VQA (76.9 vs. 77.4), where the label is known. It does not transfer, as a source label is not an intrinsic image property. Learned routers confirm this: LoRAMoE and MoLE reach 34.40 and 35.14, below the same experts routed by GSD bucket (37.89).

Prediction 2: within one source the optimum should not move (App. G.3, lines 2133-2183; Table 18). We degrade the same 2,000 RSVQA-HR pairs from 0.15 to 2.4 m effective GSD; token count is fixed and nothing is retrained. The preferred gate setting moves from 0.15 to 1.50 m, one decade. Giving the gate the effective GSD instead of the native one adds 0.9, 3.3, 5.6 and 7.6 points at 0.30, 0.60, 1.2 and 2.4 m. SSE-U, from pixels alone, tracks the effective GSD with a log-log slope of 0.91. We agree the shuffle alone could not show this, as App. A.5 (lines 1062-1063) now clarifies.

W4. Does the headline evidence identify the adapter? We agree the matched controls you highlight as "much more relevant" are the core evidence for the adapter: from the same Stage-1 checkpoint, corpus and steps, CS-HLoRA reaches 40.71 on OmniEarth-Bench vs. 31.98 for standard LoRA and 37.89 for GSD-bucketed experts (Table 2, lines 815-824), alongside the corpus-by-adapter factorial and shuffle (Table 6, lines 1028-1035). Table 1 is a system-level result, so we revised the abstract (lines 029-030) and Sec. 4.3 (lines 496-504) to lead directly with the matched numbers.

Taken together, the three tests you proposed consistently support GSD-specific gating over a generic conditional adapter.

Reviewer_Y69COpenReview ↗

Discussion

Thank you for the additional experiments. The alternative conditioning controls, within-image resolution intervention, and rank-group ablations directly address several of my concerns. These results strengthen the empirical evidence for GSD-conditioned adaptation.

My remaining concern is the interpretation of rank specialization. Since the rank groups are already assigned GSD-dependent activation patterns, the observed task-specific performance drops may partly reflect differences in gate utilization rather than independently learned object-, structure-, and semantic-level functions.

Could the authors report the group-by-task ablation results within matched GSD ranges, ideally controlling for the activation magnitude or update norm of each group? This would help establish whether the claimed functional specialization persists beyond the imposed gating structure.

I also appreciate the clarification that sampling physics motivates rather than derives the gating mechanism. The new evidence supports GSD-specific conditioning, but I remain less convinced that it establishes a physically determined hierarchy of adaptation subspaces. This distinction remains central to my assessment.

Thank you for taking the time to revisit our response and for your thoughtful follow-up questions. Your comments have helped us clarify how to distinguish learned specialization from the effects of gate activation and update magnitude. We conducted the requested ablation and report the results below.

Matched GSD range. We repeat the group-by-task removal inside the Toronto subset alone (0.05–0.1 m, N=1,240). In this range the gates of all three groups are above 0.99 for every image, except one object rank (Appendix G.7). All three groups therefore operate with nearly saturated gates.

Matched update norm. On Toronto the object, structure and semantic groups carry 19%, 33% and 48% of the gate-weighted update norm ∥hkB:,kAk,:x∥\lVert h_k B_{:,k}A_{k,:}x\rVert. We attenuate the structure and semantic gates until the removed norm equals that of the whole object group, and we match the random-rank control in norm as well.

Accuracy loss in points on object / structure / scene task families:

RemovedFull removalEqual removed norm
object group8.4 / 1.5 / 0.88.4 / 1.5 / 0.8 (same by construction)
structure group1.7 / 5.4 / 1.81.0 / 3.4 / 1.1
semantic group2.5 / 2.8 / 7.11.1 / 1.2 / 3.6
11 random ranks1.2 / 1.3 / 1.21.3 / 1.2 / 1.2 (norm-matched)
21 random ranks2.1 / 2.0 / 2.2
32 random ranks3.2 / 3.0 / 3.1

The selectivity persists. With full removal, each group lowers its own task family by 8.4, 5.4 and 7.1 points, against 1.2, 2.0 and 3.1 for the same number of random ranks. At equal removed update norm, the corresponding losses remain 2.8 to 6.5 times those of the matched random control. Losses on the other task families are at or below their controls, except for object-group removal on structure tasks (1.5 versus 1.2–1.3). This pattern persists within the matched GSD range, where all three groups have nearly saturated gates.

On the physical basis of the hierarchy. The revised manuscript explains the physical rationale in lines 37–42 and 314–323, the training mechanism in lines 486–491, and the controlled resolution experiment in Appendix G.3 (Table 18; lines 2136–2159). We appreciate your emphasis on the basis of the hierarchy and would like to make the connection between these points clearer.

As GSD increases, a fixed-size object occupies fewer pixels. Fine object cues can therefore disappear while broader spatial patterns remain observable. At finer resolution, both kinds of evidence are available. This motivates a cumulative organization: groups supporting broader context remain active, while additional ranks contribute when finer detail can be resolved. The nested gates encode this relationship between spatial sampling and the evidence available for adaptation.

Equation (1) makes this relationship part of learning. The gate weights each rank's contribution, while AA and BB learn what it computes. Ranks with lower thresholds receive stronger gradient weights from finer imagery, whereas ranks with higher thresholds are trained across a wider range of scales. The rank directions and thresholds remain learnable, but physical scale shapes which examples train each part of the adapter. Consistent with this mechanism, fixing the gate at 1.0 m during inference retains a 2.8-point advantage over standard LoRA on XLRS-Bench-lite (lines 486–491). This supports a lasting effect of scale-dependent training on the learned weights.

The resolution intervention tests whether this organization responds to available detail. In Table 18, the best gate setting shifts from 0.15 m to 1.50 m as effective GSD increases from 0.15 m to 2.4 m, with scene, sensor, question and model weights fixed. At the coarsest setting, using effective GSD improves accuracy by 7.6 points over retaining the native-GSD gate. The preferred configuration thus follows the loss of image detail even within the same source. The new Toronto ablation complements this result: the groups retain their task selectivity within a matched GSD range and at equal removed update norm. Together, these findings connect the sampling constraint, the training mechanism and the groups' learned functions. This is why we interpret CS-HLoRA as a physically motivated hierarchy of adaptation subspaces whose every link is supported by an intervention.

Thank you again for the careful and constructive discussion. Each experiment you proposed has made the paper more precise, and we are grateful for the time you have invested in it. If the new results address your remaining concern, we would sincerely appreciate it if you would consider raising your score.

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

The paper proposes ScaleEarth, a scale-conditioned adaptation framework for remote-sensing vision–language models. Following domain-specific supervised fine-tuning, it freezes the backbone and introduces CS-HLoRA, which uses continuous log-GSD to gate the rank dimensions of a shared LoRA adapter. A complementary head, SSE-U, estimates GSD and uncertainty when metadata are unavailable, enabling an uncertainty-based fallback. The paper also constructs GeoScale-VQA, a corpus of approximately 1.5 million question–answer pairs with scale annotations and scale-aware question generation. Evaluation uses Qwen3-VL-8B on XLRS-Bench and OmniEarth-Bench, supplemented by matched adaptation baselines, component ablations, GSD perturbation tests, and quantization experiments. The reported results show improvements over standard LoRA, textual GSD conditioning, and alternative gated adaptation methods.

Critical strengths

1.The strongest empirical evidence comes from the matched-backbone comparisons. Starting from the same Stage-1 checkpoint and using the same Stage-2 corpus, ScaleEarth achieves 40.71 on OmniEarth-Bench, compared with 31.98 for standard LoRA, 32.60 for textual GSD conditioning, and 37.89 for bucketed MoE-LoRA.

2.The paper addresses a consequential property of remote-sensing imagery: physical resolution affects both the available visual evidence and the questions that can be answered reliably. Conditioning a shared LoRA adapter directly on continuous GSD offers a simple and interpretable way to incorporate this information into adaptation. This is a meaningful methodological contribution because it connects a physically relevant variable to parameter modulation without requiring separate adapters for different resolution regimes.

Critical weaknesses
  1. The scale-stratified evaluation in Tables 4–5 confounds GSD with dataset source, sensor characteristics, and task distribution. Global GSD shuffling also breaks these associations, so it does not isolate physical scale. Without controlled within-source scale variation or held-out-source evaluation, the gains cannot be confidently attributed to physical-scale generalization.

  2. The comparison with bucketed MoE-LoRA matches total parameters but changes per-example effective capacity and cross-scale parameter sharing. B4 activates one lower-rank expert, whereas CS-HLoRA can use nearly all 64 rank dimensions at fine resolutions. The results therefore support the overall design but do not isolate continuous gating as the source of improvement.

Questions for authors

1.Do you have within-source scale-controlled or held-out-source results that help distinguish physical-scale adaptation from source-specific specialization? If so, do the gains persist under these controls?

1 comment
AuthorsOpenReview ↗

Response to Reviewer okSE

Thank you for the time and care you have devoted to reviewing our manuscript. We have revised the paper to address the issues you raised and describe the changes and their rationale below. We hope these revisions resolve your concerns and would appreciate your reassessment of the revised manuscript, including consideration of a higher rating if you find that the concerns have been adequately addressed.

W1 / Q1: physical-scale adaptation or source-specific specialization?

The source-scale confound is now stated in Sec. 4.2 (lines 429-431), Appendix A.3 (lines 943-947) and the shuffle discussion in Appendix A.5 (lines 1061-1063). A within-source control is reported in Sec. 4.4 (lines 494-499) and in full in Appendix G.3 (lines 2135-2185; Table 18 and Fig. 19).

The paper acknowledges that each bin in Tables 4-5 corresponds to a different source and that a global GSD shuffle breaks associations with source and sensor as well as scale. Those tests alone cannot distinguish the explanations you raise.

The new experiment degrades the same 2,000 RSVQA-HR test QA pairs to five resolutions while holding the underlying scene, sensor, question, model weights and token count fixed. Table 18 reports the results: the preferred gate setting moves from 0.15 to 1.50 m, and the gains over standard LoRA remain 9.5-12.3 points with effective GSD and 9.1-11.2 with the SSE-U estimate. This provides the within-source evidence missing from the original evaluation and supports adaptation to the loss of resolvable detail.

Table 6 provides complementary evidence from an external training corpus: on FIT-RS, CS-HLoRA improves over standard LoRA by 5.5 / 3.0 points on OmniEarth-Bench / XLRS-Bench, close to the 6.2 / 3.4 gains on GeoScale-VQA.

W2: does the B4 comparison isolate continuous gating?

The B4 comparison is qualified in Appendix A.1 (lines 884-887) and its interpretation is restated in Appendix A.8 (lines 1315-1318). The nested hard-gated control is reported in Appendix G.4 (lines 2187-2233; Table 19 and Fig. 20) and listed next to B4 in Table 7 (lines 1081-1091).

B4 changes gate continuity, parameter sharing and active rank together. The new control retains CS-HLoRA's shared parameterization and nested rank groups, replacing the sigmoid with a step under the same training setup.

Table 19 separates the 2.82-point OmniEarth gap into 1.42 points from nesting, sharing and active rank, followed by 1.40 points from smoothing the gate. The corresponding differences on XLRS-Bench-lite are 0.8 and 1.1 points. Smoothing also reduces the accuracy change across the 0.9-1.1 m interval from about 8 to 2.6 points. The revised interpretation attributes about half of the OmniEarth gap to smooth gating; the remaining gain belongs jointly to nesting, sharing and active rank, which this control does not separate further.

Reviewer_2F5yOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ScaleEarth, a scale-aware vision-language model for remote sensing imagery. It addresses the limitation of existing models that overlook how physical scale changes visual information and the questions an image can support. ScaleEarth uses ground sampling distance to guide low-rank adaptation and estimates scale with uncertainty when metadata is unavailable. The paper also presents GeoScale-VQA, a dataset designed to align question–answer supervision with image scale. Experiments demonstrate improved performance, particularly on scale-sensitive tasks, while maintaining efficient adaptation and deployment.

Critical strengths

1.The paper addresses a meaningful challenge in remote sensing: physical scale changes both the available visual evidence and the questions an image can support. Conditioning a shared low-rank adapter on GSD provides a principled way to incorporate this information into model adaptation.

2.Beyond benchmark comparisons, the paper includes controlled experiments separating the effects of training data and adaptation, along with shuffled-GSD tests and alternative conditioning strategies. These analyses strengthen the evidence that scale-dependent adaptation contributes to the improvements.

Critical weaknesses

1.Inaccurate GSD predictions may weaken adaptation, while fallback to a default scale sacrifices scale-specific benefits.

2.Automatically generated, scale-conditioned questions may favor the proposed method and introduce annotation errors. The dataset lacks manual verification, and the data generated by Claude or AI may lead to a large number of illusions.

3.Experiments focus on a single backbone family, leaving generalization to other architectures unclear.

4.CS-HLoRA adds GSD-dependent gates to standard LoRA, making the innovation primarily a domain-specific adaptation of existing ideas.

5.The framework effectively combines scale-aware data and adaptation, but the main factor contributing to this is the increase in data volume, which needs to be clearly understood.

Questions for authors

1.What is the most fundamental difference between CS-HLoRA and standard LoRA?
2.Why not directly include GSD in the prompt?
3.Is the improvement solely due to training on an additional data points?
4.When no metadata is available, is GSD guessed? What happens if the guess is wrong?
5.The gating works effectively around 0.15 meters, but why does the curve drop sharply for extremely small GSD values?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2F5y

Thank you for your careful review and constructive comments. We have revised the manuscript in response to your suggestions and explain the changes in our point-by-point responses below. If these revisions and clarifications address your concerns, we would be grateful if you could reassess the revised manuscript and consider a more favorable rating.

W1 / Q4: missing or inaccurate GSD

We address this in Appendix A.8 (lines 1318-1323) and with the estimation-error comparison in Appendix G.3 (lines 2135-2185; Table 18).

SSE-U predicts log-GSD and uncertainty from image features, falling back to 1.0 m when predicted standard deviation is at least 0.3. Benchmark evaluations already use this metadata-free procedure. Its log-GSD MAE is 0.12 (about a 1.3-fold error scale); fallback occurs on 8.4% of samples, and disabling it costs 0.7 points. Using 1.0 m for every image costs 1.4 points on XLRS-Bench-lite. In Table 18, estimated-GSD accuracy stays within 1.3 points of effective-GSD accuracy across five resolutions. Large errors can still reduce accuracy.

W2: generated QA quality and bias

We report an audit in Sec. 2.4 (lines 217-220) and Appendix G.6 (lines 2256-2275), with residual errors acknowledged in Appendix F (lines 1986-1990).

For 82% of the corpus, existing QA is unchanged; we only attach GSD. Qwen3-VL-32B generates the remaining 18%, with Claude used only as a filtering judge. Two annotators audited 500 accepted pairs: 96.4% were correct and grounded, 2.2% unsupported and 1.4% ambiguous (κ=0.83\kappa=0.83). All controlled adapters share the training corpus. On external FIT-RS, which contains none of our generated QA, CS-HLoRA still gains 5.5 / 3.0 points over LoRA (Table 6), showing that the gains do not require our generated questions.

W3: a second backbone family

We report InternVL3-8B in Appendix G.5 (lines 2235-2254; Table 20), with the scope stated in Appendix F (lines 1995-1997).

Both adapters start from the same public InternVL3-8B checkpoint without Stage-1 SFT, using 300K samples and three seeds. CS-HLoRA gains 5.5 / 3.0 points on OmniEarth-Bench / XLRS-Bench, close to 6.2 / 3.4 on Qwen3-VL-8B. Comparisons are within each backbone; evidence now covers two families, not arbitrary architectures.

W4 / Q1: difference from LoRA and contribution

We clarify this in Sec. 3.2 (lines 317-319) and Appendix E (lines 1961-1966), and add a control in Appendix G.4 (lines 2187-2233; Table 19).

CS-HLoRA replaces LoRA's shared BABA update with B diag(h(s))AB\,\mathrm{diag}(h(s))A, where s=log⁡10(GSD)s=\log_{10}(\mathrm{GSD}). GSD routes gradients during training and reweights nested ranks at inference. Fixing the trained gate at 1.0 m retains a 2.8-point advantage over LoRA on XLRS-Bench-lite; per-image gating adds 1.4 (Sec. 4.3, lines 476-482). Matched learned routers, LoRAMoE/MoLE, score 34.40/35.14 on OmniEarth versus 40.71 for CS-HLoRA (Table 2). Table 19 isolates a 1.40-point gain from smoothing a matched nested hard gate. The contribution is continuous GSD conditioning of one shared adapter, supported empirically rather than derived from physics.

W5 / Q3: additional data volume

The data-volume controls are summarized in Sec. 4.4 (lines 507-509) and detailed in Appendix A.5 (Table 6; shuffle control in lines 1058-1063).

With checkpoint, corpus and steps fixed, CS-HLoRA scores 40.71 versus LoRA's 31.98 on OmniEarth-Bench (Table 2). At 300K samples per cell and three seeds, Table 6 shows adapter gains of 5.5 / 3.0 points on FIT-RS, versus corpus gains of 2.8 / 1.5 with LoRA. CS-HLoRA trained on 300K external samples also beats LoRA trained on 1.49M GeoScale-VQA samples (34.5 versus 31.98). More data alone does not explain the gain. Shuffling GSD removes the gain, though it also breaks source associations.

Q2: GSD in the prompt

This comparison is in Appendix A.8 (lines 1290-1298) and Table 7 (lines 1081-1091).

Under matched training, GSD prompting adds 0.62 points over LoRA, versus 8.73 for CS-HLoRA. In Fig. 10, prompting shows no clear scale-dependent response (74.7-78.9%), whereas CS-HLoRA peaks near the true GSD at about 86%. Textual GSD does not reproduce direct adapter conditioning.

Q5: the small-GSD drop

We explain this in Appendix A.8 (lines 1300-1310) and with an intervention in Appendix G.7 (lines 2277-2306; Fig. 21).

Fig. 10 varies supplied GSD on images whose true GSD remains 0.15 m. Supplying 0.01 m activates a weakly trained rank with a learned cut-off of 0.039 m, below the finest training GSD (0.062 m). Restoring only this rank's gate raises accuracy from 70% to 84%, versus 86% at the true GSD: 87.5% of the loss is recovered. This explains most of the mismatch penalty; the curve does not measure performance on genuinely 0.01 m imagery.

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to address my concerns. The new results have largely resolved my questions regarding GSD estimation, data quality, backbone generalization, and the sources of performance gains. While I still have some reservations about the methodological novelty, I acknowledge that the additional evidence has substantially strengthened the paper. I am therefore willing to raise my rating.

AuthorsOpenReview ↗

Thank you for your thoughtful review and constructive feedback

Thank you for carefully reading our responses and for reconsidering your assessment. Your questions helped us identify where the paper needed stronger evidence. Working through these concerns led to experiments that improved the paper and clarified our own understanding of the method. We are glad that the results have addressed most of your questions, and we appreciate the time and care you have devoted to this discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank you to all reviewers for the constructive discussion

Thank you all for the time and care you have devoted to our paper throughout the review process. We appreciate your close reading of the manuscript and your engagement with our responses and additional results. Reviewing a paper and returning to assess its revisions takes considerable effort, and we are grateful for the attention each of you has given our work.

Your feedback helped us recognize where our explanations needed to be clearer and where stronger evidence was needed. Working through your comments led us to revisit parts of the analysis and strengthen the experimental support for our conclusions. It also helped us better understand our own method. We have learned from this exchange, and the paper has improved through the questions and suggestions you raised.

Thank you again for the constructive discussion and for considering the work we put into the revisions. We are grateful to all three reviewers for helping us improve the paper and communicate its contribution more clearly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.