- 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
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 . 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:
| Removed | Full removal | Equal removed norm |
|---|---|---|
| object group | 8.4 / 1.5 / 0.8 | 8.4 / 1.5 / 0.8 (same by construction) |
| structure group | 1.7 / 5.4 / 1.8 | 1.0 / 3.4 / 1.1 |
| semantic group | 2.5 / 2.8 / 7.1 | 1.1 / 1.2 / 3.6 |
| 11 random ranks | 1.2 / 1.3 / 1.2 | 1.3 / 1.2 / 1.2 (norm-matched) |
| 21 random ranks | 2.1 / 2.0 / 2.2 | |
| 32 random ranks | 3.2 / 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 and 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.
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:
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.