acceptodds
Under review as a conference paper at ICLR 2027

Domain Generalization Guided by Class-wise Domain-Invariant Semantic Subspaces

Abstract

Domain generalization aims to learn a model from multiple source domains that generalizes to unseen target domains. Existing domain generalization methods typically model each class as a single prototype vector, that is, a point in Euclidean space, making it difficult to explicitly characterize intra-class variation under domain shift at the class level. To this end, we model each class as a semantic subspace. Specifically, we use multiple class-related prompts to estimate the class semantic center and its principal variation directions, extending the point representation to a low-dimensional affine subspace. The model combines prototype similarity with subspace projection similarity, thereby providing semantic discriminative evidence beyond a single-point prototype for samples of the same class across different domains. To preserve the semantic structure of the subspace and improve prediction stability against variations not represented by it, we combine semantic geometry constraints with residual-space robustness learning. Theoretical analysis characterizes the representational properties of the subspace and prediction stability under bounded residual perturbations. Experiments on multiple domain generalization benchmarks show that our method achieves stable and competitive performance.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Domain Generalization, Subspace, VLMs

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3kBROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

This paper addresses domain generalization in vision-language models. It argues that standard single-text prototype class representations in CLIP are limited under domain shift. To solve this, the authors propose representing each class as a low-dimensional semantic subspace and classifying images by measuring their alignment with both the class centroid and its subspace. Across five DomainBed benchmarks, the approach achieves consistent improvements over CLIP prompt-learning baselines. Aablation study shows that the performance gain comes from the subspace structure itself, rather than simply using multiple prompts.

Critical strengths
  1. The idea is simple and easy to follow. Instead of collapsing all the text descriptions of a class into a single prototype, the paper keeps their spread as a low-rank subspace, and scores an image by both its distance to the class center and its fit to the subspace. The write-up is clear, and the method looks straightforward to reproduce.
  2. Tables 3 and 4 provide controlled comparisons that isolate the effect of the subspace structure. With the same 30 descriptions per class, averaging or taking the max over prompts performs worse than organizing them into a subspace (Table 3), and replacing each class's basis with another class's basis reduces accuracy (Table 4). These results indicate that the improvement is tied to the subspace structure and to class-specific directions, rather than to the use of multiple prompts alone.
Critical weaknesses
  1. The related work is incomplete and the main concern is that the core idea is not new. The paper situates itself only within CLIP-based prompt learning, and does not discuss the line of DG research that likewise models classes or features through subspaces and variation directions ([1] ISR, ICML 2022; [2] FVD, ACM MM 2021), nor recent CLIP-side work using diverse text prompts for DG ([3] Wen et al., CVPR 2025). Since the idea of "a class is a set of variation directions rather than a point" has precedents, a discussion of these works seems necessary to clarify what is new here.
  2. Justification of key design choices is thin. r=8r=8 is chosen without showing the singular-value spectrum (e.g., cumulative energy), which is a natural and expected analysis for an SVD-based method. In addition, the loss weights (λres=0.01λ_{res}=0.01, λinv=0.1λ_{inv}=0.1) and the adversarial radius εε have no reported sensitivity analysis, while Table 2 shows each auxiliary objective contributes within roughly one point, making it difficult to judge how stable these contributions are.
  3. The method depends on descriptions from a single LLM. The prompt bank comes entirely from GPT-5.6, and that is the only source ever tested. Since the whole method is built on top of these descriptions, I wonder how much the generator matters: would descriptions from a smaller or open-source LLM give similar subspaces and results? And what happens if some descriptions are wrong or repetitive? At this point it is hard to tell whether the gains come from the general idea of using multiple descriptions, or specifically from what GPT-5.6 knows. There is also a practical reproducibility concern, since GPT-5.6 is proprietary and its outputs may change over time — releasing the full prompt bank helps, but the pipeline itself cannot be re-run independently

[1] Provable Domain Generalization via Invariant-Feature Subspace Recovery. ICML 2022

[2] Domain Generalization via Feature Variation Decorrelation. ACM MM 2021

[3] Domain Generalization in CLIP via Learning with Diverse Text Prompts. CVPR 2025

Questions for authors
  1. What do the per-class singular-value spectra look like, and do they support r=8? How does performance change with λ_res, λ_inv (e.g., ±10×), and ε?
  2. How sensitive is the method to the description source? Would descriptions from a different LLM or human-written attributes yield similar subspaces and performance?
  3. The paper's central claim is that the textual subspaces support domain-invariant prediction, but I could not find direct evidence for this, e.g., a visualization or measurement showing that unseen-domain features lie closer to the class subspaces than a point-prototype baseline. Is there a way to demonstrate this, even on a toy setup?
2 comments

W1

Thank you for pointing out these omissions. We agree that subspace-based feature modeling has precedents. We will discuss ISR, FVD, and Wen et al., and position our contribution as class-description-derived affine geometry for classification and orthogonal prediction-consistency training.

ISR recovers invariant-feature subspaces from first- or second-order class-conditional statistics across training environments, with guarantees under its data model. We instead estimate class centers and bases from textual descriptions for image–class matching and local perturbation constraints. Our objective therefore differs from invariant-feature recovery.

FVD separates variation in visual feature space and applies class-decorrelation regularization to mitigate negative transfer. We retain principal textual variation directions as part of class representations, use affine projections for classification, and constrain predictions under bounded orthogonal perturbations. The direction sources, classification roles, and training constraints differ.

Wen et al. use domain-simulating prompts for feature suppression, consistency, and diversification. Our class-specific textual geometry instead jointly defines the affine classifier and orthogonal consistency perturbations, connecting representation and stability training through the same bases.

Controlled comparisons support these designs. With identical descriptions and a shared training objective in Table 3, our subspace outperforms maximum similarity by 1.11.1, 0.90.9, and 0.90.9 percentage points on TerraIncognita, VLCS, and OfficeHome. The additional same-rank linear control achieves 84.27±0.18%84.27 \pm 0.18\% on OfficeHome, versus 85.2±0.1%85.2 \pm 0.1\% for the affine construction, a gain of 0.930.93 percentage points. Table 4 retains class centers and rank while shuffling class-to-basis correspondence during training and inference, reducing VLCS and OfficeHome accuracy by 3.393.39 and 2.452.45 percentage points. We will clarify how these controls support our incremental contributions relative to prior subspace-based approaches.

W2/Q1

Thank you for identifying the insufficient justification of rank selection and hyperparameter sensitivity. We supplemented the paper with class-wise textual singular-value spectra and sensitivity analyses of the two auxiliary loss weights and adversarial perturbation radius on OfficeHome.

For rank selection, we computed the singular-value spectrum of centered description embeddings from the frozen CLIP text encoder for each of the 6565 OfficeHome classes. Cumulative energy is defined as the sum of squared leading singular values divided by the sum of all squared singular values. With 3030 descriptions per class, the first 88 directions retain a median energy of 69.54%69.54\%, with an interquartile range of 68.71%68.71\%–70.88%70.88\% and a class-wise range of 66.08%66.08\%–73.84%73.84\%. Thus, rank r=8r=8 retains a relatively consistent proportion of the principal textual variation across classes while truncating the remaining directions.

Energy retention alone does not determine classification performance. We therefore interpret these spectra alongside the existing rank ablation: increasing rank retains more textual variation but does not necessarily improve generalization. Together, the classification results and spectral analysis support r=8r=8 as an effective empirical trade-off. We will include the per-class spectra and cumulative-energy plots in the revised manuscript.

For hyperparameter sensitivity, we conducted additional single-run analyses on OfficeHome, changing one parameter at a time while keeping all other settings fixed. We tested one-tenth, the default, and ten times the default values for λres\lambda_{\mathrm{res}} and λinv\lambda_{\mathrm{inv}}, and half, the default, and twice the default radius ϵ\epsilon.

ParameterLower value: accuracyDefault value: accuracyHigher value: accuracy
λres\lambda_{\mathrm{res}}0.0010.001: 85.6%85.6\%0.010.01: 86.2%86.2\%0.10.1: 85.8%85.8\%
λinv\lambda_{\mathrm{inv}}0.010.01: 86.0%86.0\%0.10.1: 86.2%86.2\%11: 85.7%85.7\%
ϵ\epsilon0.0250.025: 85.4%85.4\%0.050.05: 86.2%86.2\%0.10.1: 85.7%85.7\%

Results are single-run classification accuracies. The accuracy ranges are 0.60.6 and 0.50.5 percentage points for the two loss weights and 0.80.8 points for the perturbation radius. The default configuration performs best among the tested values, with limited degradation elsewhere. These results indicate bounded performance variation within the tested ranges, while showing that regularization strength and perturbation radius still require an appropriate balance.

The ablations in Table 2 assess whether each auxiliary objective contributes, whereas these additional analyses examine sensitivity to their weights and the perturbation radius. We will present both in the revised manuscript to provide a more complete justification of the design choices.

W3/Q2

Thank you for raising concerns about description sources and quality. We conducted additional cross-generator experiments and controlled tests with repeated and incorrect descriptions on OfficeHome. All settings use 3030 descriptions per class, rank 88, and the same training and evaluation protocol.

We independently generated description banks using DeepSeek-V4.1-Flash-1017 and GPT-6.1. For description quality, the duplicate setting replaces six descriptions per class with copies of remaining same-class descriptions; the incorrect setting replaces them with descriptions from other classes. Replacement rules and the random seed are fixed.

Description bankAccuracy
Original GPT-5.686.2±0.1%86.2 \pm 0.1\%
DeepSeek-V4.1-Flash-101785.5±0.3%85.5 \pm 0.3\%
GPT-6.185.9±0.3%85.9 \pm 0.3\%
Original bank, 20%20\% duplicate descriptions85.08±0.3%85.08 \pm 0.3\%
Original bank, 20%20\% incorrect-class descriptions84.58±0.2%84.58 \pm 0.2\%

Changing generators reduces accuracy by 0.70.7 and 0.30.3 percentage points, respectively. Performance remains relatively close across the tested sources, indicating that the method is not effective only with the original generator. This does not imply identical subspace geometry across generators.

Duplicate and incorrect descriptions reduce accuracy by 1.121.12 and 1.621.62 percentage points, respectively. Duplicates reduce diversity and change the statistical weighting of descriptions, whereas incorrect descriptions introduce class-mismatched information. These results indicate some tolerance to the tested corruption, while confirming the importance of description quality. The incorrect setting is synthetic wrong-class contamination, not an estimate of the error rate in naturally generated descriptions.

In practice, errors and redundancy can be reduced through explicit generation instructions requiring correct class identity, observable attributes, and distinct descriptions; duplicate and semantic-similarity checks followed by regeneration; and independent-model cross-checks of class relevance and attribute plausibility, with manual inspection of questionable descriptions. Replacements should preserve the description count per class. These measures reduce risk rather than guarantee error-free banks.

We will include these findings and practical recommendations in the revised manuscript and clarify the method's scope with respect to description sources and quality.

Q3

Thank you for suggesting direct evidence of unseen-domain alignment. We conducted inference-time diagnostics on all four held-out OfficeHome domains. For each domain, we fixed the trained model, image features, and class centers, comparing correct rank-88 bases with shuffled bases of the same rank. Shuffled results are averaged over 5050 permutations.

We measure squared distance to the ground-truth class affine subspace and the classification margin, defined as the ground-truth score minus the highest incorrect-class score. Lower distance and higher margin are preferable.

Target domainCorrect-basis distance ↓Shuffled-basis distance ↓Correct-basis margin ↑Shuffled-basis margin ↑
Art1.02751.21710.12610.1042
Clipart1.06861.09040.11120.0867
Product1.03561.07720.15940.1381
Real World0.91920.97830.18580.1621
Equally weighted mean1.01271.09080.14560.1228

Correct bases consistently yield smaller distances and larger margins, supporting both geometric alignment and discriminative value on unseen domains.

For the point-prototype comparison, the class center lies within its affine subspace, so subspace distance cannot exceed center distance. Smaller distance alone therefore does not establish the usefulness of class-specific directions. Our equal-rank, shared-center shuffled control specifically tests these directions.

The complementary training ablation in Table 5 gives 85.0±0.1%85.0 \pm 0.1\% for prototype-only classification, 85.3±0.1%85.3 \pm 0.1\% for projection-only classification, and 86.2±0.1%86.2 \pm 0.1\% for their combination. Together, these results show that subspace information supplements the prototype and that correct textual directions improve unseen-domain alignment and classification margins.

We will include these diagnostics while limiting our conclusion to their demonstrated geometric and discriminative value, rather than claiming complete feature-distribution invariance across domains.

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

The paper represents each class as a low-rank textual subspace derived from multiple prompts and combines prototype and subspace-projection similarities for classification. It uses semantic-preservation losses and adversarial consistency outside the subspace to improve robustness under domain shift. Experiments on five DomainBed benchmarks with CLIP ViT-B/16 show consistent gains over reproduced baselines, supported by ablations and geometric and stability analyses.

Critical strengths
  • The subspace representation makes better use of multiple descriptions. Rather than averaging them into one prototype, the method retains their main variation directions. The ablations show that both the class center and subspace projection contribute to classification.

  • The paper includes useful controls. Using the same 30 descriptions across methods helps separate the benefit of subspace modeling from simply providing more text. Swapping subspace bases between classes also reduces performance, supporting the importance of class-specific geometry.

  • The gains are consistent across datasets. The method outperforms the reproduced baselines on all five benchmarks, improving average accuracy from 75.3% for MaPLe to 77.9%. Results over three seeds and loss ablations provide further support.

Critical weaknesses
  • The relevance of the evaluation setting could be clearer. I have not followed this area closely in recent years, so I would appreciate more context on how challenging these benchmarks remain for current vision–language models. With the rise of multimodal LLMs such as Qwen-VL, how much of the domain gap persists with stronger pretrained models? Some evidence of the remaining headroom would help explain the practical importance of these improvements.

  • The distinction from earlier text-based approaches needs more discussion. Multiple descriptions already enrich class representations in CuPL and classification by description. ProDA models variation across prompt embeddings, while Wen et al. use diverse text prompts specifically for domain generalization. The same-description controls are useful, but compare only against averaging and maximum similarity. A closer discussion and comparison would help establish what the subspace classifier and residual-space training add.

  • The connection between textual variation and visual domain shift remains uncertain. Principal directions among generated descriptions may capture wording or attribute differences without corresponding to changes across image domains. Directions outside the textual subspace may also contain useful class information. Comparisons with subspaces built from augmented source-image features and with unrestricted feature-space adversarial consistency would help test whether the proposed geometry matters.

  • Robustness to the description bank is unclear. The description-count study supports choosing 30 descriptions, but uses nested subsets of one generated bank. Independent banks or randomly sampled subsets would better establish stability across wording choices. Please also clarify how rank is handled with five descriptions: their centered features support at most four independent directions, below the default rank of eight.

Questions for authors
  1. Relation to prior work: What distinguishes the proposed subspace representation from ProDA and DG with Diverse Text Prompts? Can you clarify which benefits are specific to the subspace formulation?

  2. Role of the residual-space restriction: Does restricting adversarial perturbations to the orthogonal complement outperform unrestricted feature-space adversarial consistency under the same training settings? This would help establish whether the textual geometry provides a useful constraint.

  3. Description-bank robustness: Are the improvements consistent across independently generated description banks or randomly sampled subsets?

4 comments

W1

Thank you for suggesting that we examine the remaining headroom on these benchmarks with more recent vision-language models. We conducted additional zero-shot classification experiments on PACS, OfficeHome, and TerraIncognita, comparing CLIP ViT-B/16 with Qwen2.5-VL-3B-Instruct. For each dataset, both models use the same fixed image samples and candidate classes, without fine-tuning on labeled images from these datasets. We report equally weighted mean accuracy across the four domains and worst-domain accuracy.

DatasetSamplesCLIP meanQwen-VL meanCLIP worst domainQwen-VL worst domain
PACS89696.99%96.99\%94.42%94.42\%91.07%91.07\%87.50%87.50\%
OfficeHome1,04079.33%79.33\%71.06%71.06\%64.23%64.23\%67.69%67.69\%
TerraIncognita1,12922.94%22.94\%39.86%39.86\%17.99%17.99\%35.29%35.29\%

The remaining difficulty varies substantially across benchmarks. Both models exceed 94%94\% mean accuracy on PACS, whereas approximately one-third of images remain misclassified in their respective worst domains on OfficeHome. TerraIncognita provides particularly direct evidence: Qwen-VL outperforms zero-shot CLIP at all four camera locations, improving mean accuracy by 16.9116.91 percentage points, yet reaches only 39.86%39.86\% mean accuracy and 35.29%35.29\% worst-location accuracy. Thus, cross-location recognition remains challenging even with a pretrained model that performs better on this benchmark.

These findings also contextualize the practical value of our results. Under the fixed CLIP ViT-B/16 backbone and source-domain adaptation protocol, our method improves TerraIncognita target-domain mean accuracy from MaPLe's 53.6%53.6\% to 58.7%58.7\%. We will report these two evaluations separately: the zero-shot experiments identify domains where the evaluated pretrained models still struggle, while the main leave-one-domain-out experiments measure our adaptation gains with a fixed backbone. The domain-wise differences in the zero-shot evaluation describe classification performance across domains; measuring the source–target accuracy gap after source-domain adaptation would require evaluation under the corresponding training protocol.

W2/Q1

Thank you for highlighting these related works. We agree that multiple-description representations and prompt-embedding variation have important precedents. We will expand the related-work discussion and clarify our representation, classifier, and training objectives.

CuPL and Classification by Description establish the value of generated descriptions and visual attributes. We organize these descriptions into class centers and low-rank variation directions, combining center and affine-projection similarities for classification. ProDA models prompt-embedding variation using Gaussian distributions to formulate its training objective; we instead estimate orthonormal bases from centered descriptions for projection-based classification and geometry-guided consistency training.

Wen et al. use domain-simulating prompts for feature suppression, consistency, and diversification. Our class-specific textual geometry instead determines both classification scores and perturbation directions, constraining the full class prediction distribution under bounded perturbations orthogonal to the ground-truth class subspace.

Controlled experiments isolate these contributions. In Table 3, all multi-description variants share the same 3030 descriptions per class, prompt adaptation, and training objective. Our subspace outperforms maximum similarity by 1.11.1, 0.90.9, and 0.90.9 percentage points on TerraIncognita, VLCS, and OfficeHome, respectively. The additional same-rank linear control achieves 84.27±0.18%84.27 \pm 0.18\% on OfficeHome, versus 85.2±0.1%85.2 \pm 0.1\% for the affine construction in Table 3.

Table 4 retains class centers, rank, and the prompt bank while shuffling class-to-basis correspondence during training and inference, reducing VLCS and OfficeHome accuracy by 3.393.39 and 2.452.45 percentage points. Table 2 further shows that removing residual prediction consistency reduces OfficeHome accuracy from 86.2%86.2\% to 85.6%85.6\%. These controls support the respective roles of description organization, class-specific geometry, and residual consistency, which we will clarify in the revised manuscript.

W3/Q2

Thank you for raising the connection between textual geometry and visual domain shift. We agree that textual directions may reflect both visual attributes and wording variation, while their orthogonal complement may retain useful class information. We therefore use textual subspaces as semantic priors for bounded consistency training; classification still scores the complete image features against all classes.

To test the orthogonal restriction, we added unrestricted feature-space consistency on OfficeHome. Both settings use the same classifier, KL objective, radius ϵ=0.05\epsilon=0.05, step size η=0.05\eta=0.05, three attack steps, and weight λinv=0.1\lambda_{\mathrm{inv}}=0.1. Only the perturbation-direction restriction differs.

The first two rows reproduce the setting without consistency and the full model from Table 2; the third is the additional control. Results are mean accuracy and standard deviation across three seeds.

Training settingAccuracy
Without consistency, Table 285.6±0.1%85.6 \pm 0.1\%
Orthogonal residual consistency, full model86.2±0.1%\mathbf{86.2 \pm 0.1\%}
Unrestricted consistency, additional control84.35±0.48%84.35 \pm 0.48\%

Orthogonal residual consistency improves accuracy by 0.600.60 percentage points over removing consistency and 1.851.85 points over unrestricted consistency. This supports the direction constraint under the evaluated training settings and perturbation budget. Unrestricted perturbations may alter class-discriminative directions, making consistency against such changes detrimental; our results are consistent with this explanation.

Subspaces from augmented source-image features are also worth investigating. Unlike text-derived bases, their geometry depends on source-image sampling and augmentation. Uneven class frequencies and domain variation may bias pooled directions toward particular domains or backgrounds, requiring decisions about class/domain balancing and which augmentation-induced variations should define the subspace.

Basis updates introduce further choices because visual features adapt through learnable prompts. Fixed initial bases may become misaligned with evolving features, whereas dynamic bases require specifying update frequency, sample selection, and the treatment of historical features. These choices affect stability and computational cost.

We have not completed a systematic comparison of these image-based constructions and cannot establish the relative advantages of the two basis sources. We will clarify this scope and identify source-image-guided and joint image–text subspaces as future directions.

W4/Q3

Thank you for distinguishing description count from description selection. Appendix A.3 uses nested subsets of a single fixed bank to examine the effect of description count. To further assess sensitivity to description sources, we independently generated 3030 descriptions per class using DeepSeek-V4.1-Flash-1017 and GPT-6.1, and evaluated our method under the same OfficeHome settings.

The two new banks achieve 85.5±0.3%85.5 \pm 0.3\% and 85.9±0.3%85.9 \pm 0.3\%, respectively, compared with 86.2±0.1%86.2 \pm 0.1\% using the original GPT-5.6 bank. The decreases of 0.70.7 and 0.30.3 percentage points indicate reasonably stable performance across the tested sources, while confirming that description sources still affect accuracy. We will include these results and release the corresponding banks.

For the five-description setting, Appendix A.3 uses the effective rank $

rK=min⁡(r,K−1,D),r_K=\min(r,K-1,D),

$ where KK is the number of descriptions per class and DD is the text-embedding dimension. The main experiments use K=30K=30 and rank 88, whereas the K=5K=5 setting uses rank 44, since five centered embeddings support at most four independent variation directions. We will explicitly state the effective ranks in the caption of Figure 4. All additional description-source experiments use K=30K=30 and rank 88, isolating bank changes while keeping description count and subspace dimension fixed.

Thanks for the detailed rebuttal. The Qwen-VL zero-shot results address my headroom question (W1). The unrestricted-consistency control supports the orthogonal restriction, and the independent description banks and rank clarification resolve W4.

My main concerns remain. The distinction from ProDA and Wen et al. (W2) is argued in text but not tested in experiments. Image-feature subspaces were not compared (W3), so it's still unclear whether textual directions reflect visual domain shift. I'll keep my score at 2.

W2

Thank you for clarifying that our methodological discussion required direct experimental comparisons. We have now added ProDA and Wen et al. on OfficeHome and PACS.

ProDA's code download link returns “404 Not Found,” and we found no public code link for Wen et al. We therefore independently implemented both methods from their published formulations, verifying the core computations and retaining their respective prompt and training mechanisms. Experiments use CLIP ViT-B/16 with the manuscript's data splits and model-selection protocol. We report mean accuracy and standard deviation across two seeds.

MethodOfficeHomePACS
ProDA, our reimplementation84.55±0.3%84.55 \pm 0.3\%96.68±0.3%96.68 \pm 0.3\%
Wen et al., our reimplementation85.43±0.4%85.43 \pm 0.4\%96.52±0.2%96.52 \pm 0.2\%
Our method86.2±0.1%\mathbf{86.2 \pm 0.1\%}97.6±0.1%\mathbf{97.6 \pm 0.1\%}

Our method improves over ProDA by approximately 1.651.65 and 0.920.92 percentage points on OfficeHome and PACS, respectively, and over Wen et al. by 0.770.77 and 1.081.08 points.

ProDA models classifier-weight distributions through learnable prompts; Wen et al. use diverse templates for feature suppression, consistency, and diversification. Our affine class geometry instead jointly supports projection-based classification and orthogonal residual consistency. These complete-method comparisons complement the same-description controls in Table 3, basis shuffling in Table 4, and the additional linear-versus-affine control, which examine specific contributions. We do not attribute the entire between-method gap to a single component. We will include these results, label them as independent reimplementations, and document implementation choices and differences from the original settings.

W3

Thank you for pointing out that the orthogonal-consistency experiment does not replace an image-feature subspace comparison. We have now added class-specific subspaces constructed from augmented source-image features on OfficeHome and PACS, following the original manuscript's data splits and model-selection protocol.

Specifically, we select class-specific images exclusively from the source-domain training sets and fix the selected images and their augmented views. During training, their features are recomputed using the current adapted visual encoder, then centered and decomposed by SVD to obtain class centers and rank-88 variation directions. Thus, the image-derived geometry evolves with visual-prompt learning rather than remaining fixed at initialization. Both class centers and bases in this control are derived from source-image features; target-domain images are not used in their construction.

The text- and image-derived constructions use the same subspace rank, classification rule, and shared training objectives to examine the role of the geometry's source. Results are mean accuracy and standard deviation across two random seeds.

Subspace constructionOfficeHomePACS
Augmented source-image features82.67±0.2%82.67 \pm 0.2\%95.12±0.3%95.12 \pm 0.3\%
Class descriptions, our method86.2±0.1%\mathbf{86.2 \pm 0.1\%}97.6±0.1%\mathbf{97.6 \pm 0.1\%}

The text-derived construction improves accuracy by approximately 3.533.53 and 2.482.48 percentage points on OfficeHome and PACS, respectively. Under the evaluated construction, textual geometry is therefore more effective for unseen-domain classification than geometry estimated from augmented source images. Principal source-image variations may also reflect source-specific appearance, backgrounds, and sampling distributions; the results are consistent with textual descriptions providing complementary semantic information.

A complementary inference-time diagnostic examines the usefulness of the textual directions on unseen domains. On all four OfficeHome target domains, we fixed the trained model, image features, and class centers, replacing each correct class basis with another class's basis of the same rank. Shuffled results were averaged over 5050 permutations. Correct bases yield a four-domain mean squared distance to the ground-truth class affine subspace of 1.01271.0127, compared with 1.09081.0908 for shuffled bases, and a mean classification margin of 0.14560.1456, compared with 0.12280.1228. All four target domains exhibit the same direction of improvement.

The image-subspace comparison tests how the geometry's source affects classification, while the shuffled-basis diagnostic shows that matching unseen-domain images to correct textual directions has discriminative value. The orthogonal-versus-unrestricted consistency comparison supports using this geometry to constrain perturbation directions.

Together, these experiments support the practical value of textual subspaces for unseen-domain prediction. We will add the direct comparisons and limit our conclusions to the evaluated constructions and settings, without claiming a strict correspondence between textual directions and every form of visual domain shift.

Reviewer_jsv2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a CLIP-based domain generalization method that represents each class as a low-dimensional affine textual semantic subspace rather than a single prototype. Multiple class-related prompts are used to estimate the class semantic center and its principal variation directions. To preserve the semantic structure and improve prediction stability, the authors combine textual knowledge and subspace geometry preservation with weak subspace-proximity regularization and adversarial residual-space prediction consistency. Experiments on five domain generalization benchmarks show competitive performance compared with the selected CLIP-based methods.

Critical strengths
  1. Representing each class with a center and a few variation directions is intuitive.
  2. Experiments cover five standard DG benchmarks.
  3. The comparisons with prompt averaging and maximum similarity use the same descriptions, which helps isolate the effect of the subspace representation.
  4. Details of prompt generation and model selection are provided in the appendix, making the experimental setup easier to follow.
Critical weaknesses
  1. The central subspace formulation has a close precedent that is not discussed. The publicly available manuscript [1] also constructs class-specific subspaces from multiple prompt embeddings using SVD and classifies images by cosine similarity to their subspace projections. The present work introduces an affine center, prototype fusion, and residual-space regularization, but provides no comparison with this closely related formulation. This makes it difficult to assess whether these extensions constitute a substantial methodological advance.
  2. The experimental comparison omits several strong CLIP-based DG methods. For example, [2] and [3] report average accuracies of 78.6% and 79.17%, respectively, using CLIP ViT-B/16 on the same five DomainBed benchmarks, compared with 77.9% here. The current comparisons demonstrate improvements over the selected baselines but leave the method’s competitiveness against stronger existing approaches insufficiently assessed.

[1] Vision-Language Subspace Prompting, arxiv24.
[2] CLIPood: Generalizing CLIP to Out-of-Distributions, ICML23.
[3] Domain Generalization in CLIP via Learning with Diverse Text Prompts, CVPR25.

Questions for authors
  1. What is the main methodological advance over SuPr?
  2. How does the proposed method compare with stronger baselines?
  3. Under what conditions does the proposed subspace construction reduce the source–target discrepancy term in Theorem B.6?
1 comment

Q1/W1

Thank you for pointing out SuPr, a closely related work that constructs class-specific subspaces from multiple prompt embeddings and classifies images using projection-based cosine similarity. We will add the relevant discussion and direct comparison to the revised manuscript.

Our main advances are a centered affine representation and prediction consistency guided by the same textual geometry. While SuPr constructs linear subspaces through the origin, we apply SVD to centered description embeddings to explicitly represent class centers and variation directions. We further restrict bounded adversarial perturbations to the orthogonal complement of the ground-truth class subspace and constrain changes in the full class prediction distribution using KL divergence. Thus, the same subspace supports both classification and stability training.

Two controlled comparisons assess these designs. Under the OfficeHome setting of Table 3, with identical descriptions, rank, prototype branch, and training objective, the linear subspace achieves 84.27±0.18%84.27 \pm 0.18\%, whereas the affine subspace achieves 85.2±0.1%85.2 \pm 0.1\%, an improvement of 0.930.93 percentage points. Table 2 further shows that removing residual prediction consistency reduces full-model accuracy from 86.2%86.2\% to 85.6%85.6\%.

Additionally, a full-method comparison based on SuPr's public implementation yields 85.14±0.19%85.14 \pm 0.19\% on OfficeHome across three seeds, compared with 86.2±0.1%86.2 \pm 0.1\% for our full method, an improvement of approximately 1.11.1 percentage points.

Q2/W2

Thank you for pointing out this omission. We agree that omitting these strong baselines made the original comparison incomplete and limited the evidence supporting our competitiveness claims. Using CLIP ViT-B/16 on the same five DomainBed benchmarks, CLIPood and Wen et al.'s Diverse Text Prompts report average accuracies of 78.6%78.6\% and 79.17%79.17\%, respectively, exceeding our 77.9%77.9\%. We will add their per-dataset results and averages to the main table, clearly identified as results reported in the original publications.

The methods use textual information differently: CLIPood adjusts classification margins using textual semantic relationships and employs a Beta moving average of model weights; Wen et al. use text prompts that imitate different domains to guide feature learning. Our method freezes the pretrained image and text encoders and constructs affine subspaces from class descriptions for both projection-based classification and orthogonal residual-space prediction consistency.

We will accordingly narrow our performance claims: our evidence supports the advantage of subspace representations over prompt averaging and maximum similarity under the same descriptions (Table 3), while our five-benchmark average remains below the published results of these stronger baselines.

Q3

Thank you for the question. Under equal class priors in the source and target domains, the discrepancy term in Theorem B.6 is determined by the following cost for same-class feature pairs: $

∥ΠyΔz∥2+max⁡{∥(I−Πy)Δz∥2−ϵ,0},\|\Pi_y\Delta z\|_2+ \max\{\|(I-\Pi_y)\Delta z\|_2-\epsilon,0\},

$ where Δz=zQ−zP\Delta z=z_Q-z_P and Πy=ByBy⊤\Pi_y=B_yB_y^\top. Thus, the discrepancy is small if there exist class-conditional couplings with small within-subspace shifts and orthogonal shifts largely covered by radius ϵ\epsilon.

More specifically, fix the feature distributions and ϵ\epsilon. If our bases reduce the class-prior-weighted expected cost under an optimal coupling for a reference subspace, the resulting discrepancy is strictly smaller, since it can further take the infimum over all class-conditional couplings. This depends on the alignment between the bases and the actual domain shifts. Affine centers cancel when taking feature differences, so centered SVD alone does not guarantee a reduction.

Residual consistency training instead targets the prediction-sensitivity term KP,ϵ(h)K_{P,\epsilon}(h). We will clarify this distinction: Theorem B.6 provides a conditional OOD risk decomposition, rather than guaranteeing that our construction reduces the discrepancy for arbitrary unseen domains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.