acceptodds
Under review as a conference paper at ICLR 2027

Learning from Variation for Open-Domain Few-Shot Class-Incremental Learning

Abstract

Few-Shot Class-Incremental Learning (FSCIL) aims to continually acquire novel classes from limited examples while retaining previously learned knowledge. However, existing approaches typically assume that class increments occur within a fixed real-world domain, overlooking domain variations in open-world environments. In this work, we introduce Open-Domain Few-Shot Class-Incremental Learning (OD-FSCIL), where incremental samples may arise from heterogeneous domains whose composition is not predefined. Such domain heterogeneity makes intra-class variations difficult to capture from few-shot observations and complicates joint recognition of all seen classes. To address these challenges, we propose Learning from Variation (LeVa), which leverages the rich intra-class variations of data-rich base classes as prior knowledge for novel classes. LeVa comprises Variation Recovery (VR) and Variation Exploitation (VE). VR transfers historical intra-class variations to novel classes and uses predicted distributional statistics to retain those compatible with each novel class. VE exploits the recovered variations to model class-specific structures around text anchors, enabling more effective joint recognition of both old and new classes. To facilitate evaluation under open-domain conditions, we further construct Hybrid-miniImageNet, a new benchmark covering diverse visual domains. Extensive experiments on Hybrid-miniImageNet, DomainNet, miniImageNet, and CIFAR100 demonstrate the effectiveness of LeVa under both open-domain and standard FSCIL settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Few-Shot Class-Incremental Learning, Open-Domain

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_tp7FOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Open-Domain FSCIL, in which the base classes come from a single real-image domain but the few-shot samples of new classes may come from mixed, unknown domains. The proposed LeVa builds on CLIP. It transfers intra-class variations learned from base classes to new classes, keeps the variations that match a predicted class distribution, and then uses them to model variation around the text anchors for recognition. The authors also build a new benchmark, Hybrid-miniImageNet, and report gains on it, on DomainNet and on standard FSCIL benchmarks.

Critical strengths
  1. Relaxing the single-domain assumption of FSCIL is practically relevant, and the new benchmark is a useful contribution. Moving these variations between classes by parallel transport on the hypersphere is an elegant and novel design.
  2. LeVa outperforms the strongest baseline on both open-domain benchmarks and also does well in the standard setting. The ablations cover the main components.
Critical weaknesses
  1. Most baselines are ResNets trained from scratch, compared against pretrained CLIP. Only one CLIP-based method (BiMC) is included. Zero-shot CLIP (85.5 AHM) and relevant pretrained-model FSCIL methods (e.g. Goswami et al., 2024; FSHIL, CVPR 2026) are missing from the main tables, so it is hard to tell how much of the gain comes from the method itself.
  2. The distribution-guided selection, the main novelty, beats random variation transfer by only 0.9 AHM.
  3. Key parts of Variation Recovery (e.g. the descriptor q(S, t), the kernel and RFF settings, the memory size) are not specified. The construction of Hybrid-miniImageNet is also unclear, including the image sources per domain and the test-set composition.
Questions for authors
  1. How exactly is Hybrid-miniImageNet built (image sources and counts per domain, test-set composition), and will it be released?
  2. Could you add zero-shot CLIP, a simple trained CLIP baseline and one or two pretrained-model FSCIL methods to Tables 1–2?
  3. How much does the learned distribution predictor help compared with simply using the support mean?
6 comments

Thank you for your thoughtful and encouraging feedback.

Response to Q1 and W3: Dataset construction and implementation details.

We thank the reviewer for pointing out these omissions and apologize for the lack of clarity in the original manuscript. We provide the missing dataset construction and implementation details below and will incorporate them into the revised manuscript.

(1) Construction and release of Hybrid-miniImageNet. Hybrid-miniImageNet preserves the original miniImageNet class split and base-class training and test data. For novel classes, Real-domain images are sampled from the original miniImageNet novel-class set, while stylized images are primarily sourced from ImageNet-R (Rendition) and ImageNet-Sketch, supplemented with web-collected images for classes insufficiently covered by these datasets. Each novel class contains 600 images across four domains—Real, Clipart, Painting, and Sketch—with 150 images per domain. We reserve 25 images per domain for testing, yielding 100 test images per novel class and 4,000 test images across all 40 novel classes. The remaining 125 images per domain constitute the training pool from which the five labeled supports per novel class are sampled. We will publicly release the dataset.

(2) Descriptor q(Se,tce)q(S_e,t_{c_e}).

We define the descriptor as

q(Se,tce)=[μ(Se);r(Se);tce],q(S_e,t_{c_e})=[\mu(S_e);r(S_e);t_{c_e}],

where semicolons denote concatenation. Here, μ(Se)\mu(S_e) summarizes the support distribution through its empirical kernel mean, r(Se)r(S_e) characterizes the pairwise visual similarity structure among support samples, and tcet_{c_e} provides the class text representation.

(3) Kernel and random Fourier features.

We use a Gaussian RBF kernel on 64-dimensional PCA projections of CLIP features. The kernel bandwidth σ\sigma is set to the median pairwise Euclidean distance between base-training features after PCA projection. We approximate the kernel using 256-dimensional random Fourier features (RFF), with frequencies sampled from N(0,σ−2I)\mathcal{N}(0,\sigma^{-2}I) and phases sampled uniformly from [0,2π)[0,2\pi). The PCA projection, kernel bandwidth, and sampled RFF parameters are determined during the base stage and remain fixed throughout subsequent incremental sessions.

(4) Variation memory.

We maintain a feature-level variation memory. Each class stores one prototype, while each variation entry stores a tangent vector linked to its corresponding class center through a class index.

On Hybrid-miniImageNet, the base memory contains 9,600 variation entries, with 160 entries per base class. Each novel class adds five entries from its observed supports. After the final session, the memory contains 9,800 variation entries and 100 prototypes, requiring approximately 20.4 MB.

Response to Q2 and W1: Comparisons with pretrained-model baselines

We thank the reviewer for highlighting the need for stronger pretrained-model comparisons. We have expanded the evaluation on both benchmarks to include zero-shot CLIP, CoOp as a trainable CLIP baseline, and two pretrained-model FSCIL methods: FSCIL-Calibration, suggested by the reviewer, and HyCal. We will incorporate these results into Tables 1–2 of the revised manuscript.

For the suggested FSHIL method, we were unable to obtain its official code and data splits at the time of our experiments and therefore do not report results for it. We instead include HyCal as an additional pretrained-model FSCIL comparison. For HyCal, we used grid search to select (α,β)=(40,5)(\alpha,\beta)=(40,5) on Hybrid-miniImageNet and (20,5)(20,5) on DomainNet.

As shown below, LeVa achieves the highest HM in every incremental session on both benchmarks. Among the added baselines, CoOp achieves the strongest AHM, demonstrating the importance of comparing against a trainable CLIP model. LeVa nevertheless exceeds CoOp by 1.84 percentage points on Hybrid-miniImageNet and 4.42 percentage points on DomainNet in AHM. Together with the controlled component ablations, these comparisons provide evidence of gains beyond simply adopting a pretrained CLIP backbone.

Table R1. Additional baseline comparisons on Hybrid-miniImageNet. S0 reports base-session accuracy (%); S1–S8 report HM (%). AHM is averaged over S1–S8.

MethodS0 Acc.S1 HMS2 HMS3 HMS4 HMS5 HMS6 HMS7 HMS8 HMAHM
Zero-shot CLIP91.8091.1087.8687.4584.4084.3682.6183.0383.2685.51
CoOp96.4094.3291.6791.2688.6288.7786.8586.6287.0789.40
FSCIL-Calibration97.8777.8370.7572.9967.8358.9758.2859.3560.4465.81
HyCal94.3889.6187.8385.9581.8681.6077.7976.9776.7582.29
LeVa96.9595.7693.9293.1390.5090.4988.5688.5289.0491.24

Table R2. Additional baseline comparisons on DomainNet. S0 reports base-session accuracy (%); S1–S8 report HM (%). AHM is averaged over S1–S8.

MethodS0 Acc.S1 HMS2 HMS3 HMS4 HMS5 HMS6 HMS7 HMS8 HMAHM
Zero-shot CLIP67.0867.0066.5665.1563.9062.4962.5562.3162.4364.05
CoOp78.7773.3271.1669.4567.4066.5367.0466.2466.5668.46
FSCIL-Calibration81.8051.5550.3146.2444.3844.1644.1541.9944.2245.87
HyCal82.7963.0558.5656.8655.7753.5952.4649.6548.7854.84
LeVa87.0178.4676.0873.7472.2570.9871.1370.1570.2272.88
  1. Zero-shot CLIP — Learning Transferable Visual Models From Natural Language Supervision [ICML 2021].
  2. CoOp — Learning to Prompt for Vision-Language Models [IJCV 2022].
  3. FSCIL-Calibration — Calibrating Higher-Order Statistics for Few-Shot Class-Incremental Learning with Pre-trained Vision Transformers [CVPR Workshops 2024].
  4. HyCal — HyCal: A Training-Free Prototype Calibration Method for Cross-Discipline Few-Shot Class-Incremental Learning [CVPR 2026]

Response to W2 and Q3: Comparisons with random selection and the support-based reference

We thank the reviewer for suggesting a closer examination of distribution-guided recovery. We acknowledge that the AHM gain over random variation selection on Hybrid-miniImageNet is modest (0.91 percentage points). We extended this comparison to DomainNet and additionally evaluated a support-based distributional reference.

(1) Selection using the learned distributional reference versus random selection. The random baseline samples from the same semantically filtered candidate pool, retains the same number of recovered features, and uses identical downstream processing. Selection guided by the learned distributional reference improves AHM by 0.91 percentage points on Hybrid-miniImageNet and 4.11 percentage points on DomainNet, with a 5.86-percentage-point gain in average novel-class accuracy on DomainNet. On Hybrid-miniImageNet, random selection already achieves 90.33% AHM, leaving relatively limited headroom for further gains. On DomainNet, which includes additional domains such as Infograph and Quickdraw, the random baseline achieves 68.77% AHM. These more challenging domain conditions may provide greater scope for the learned distributional reference to improve candidate selection.

(2) Learned versus support-based reference. We replace the predicted distributional reference μ^c\hat{\mu}_c with the empirical kernel mean of the support samples, μ(Sc)\mu(S_c), while keeping the remaining components unchanged. The learned reference improves average novel-class accuracy by 4.80 percentage points and AHM by 2.39 percentage points, with a 1.34-percentage-point decrease in average base-class accuracy. This comparison demonstrates the additional benefit of predicting a distributional reference beyond directly using the observed supports.

Table R3. Variation recovery ablations on DomainNet.

VariantAvg. Novel ↑Avg. Base ↑AHM ↑
Random selection60.2180.9868.77
Support-based reference61.2783.4270.49
Learned reference (LeVa)66.0782.0872.88

These controlled comparisons support the contributions of both distribution-guided selection and the learned distributional reference.

Reviewer_1uJwOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Open-Domain Few-Shot Class-Incremental Learning, where novel-class observations may come from heterogeneous domains whose composition is not predefined. The proposed method, LeVa, combines Variation Recovery and Variation Exploitation. It predicts a distributional reference from few-shot observations and class semantics to select compatible variations transferred from historical classes. These recovered variations are then incorporated into visual matching and text-anchored recognition for joint classification of all seen classes.

Critical strengths
  1. The paper addresses a meaningful and practically relevant extension of conventional FSCIL. The uncertainty introduced by heterogeneous domains provides clear motivation for studying how prior knowledge can support learning from limited observations.

  2. The framework follows a clear recover-and-exploit strategy. The two modules have complementary roles and provide a reasonable approach to learning from limited observations under domain heterogeneity.

  3. The evaluation covers both open-domain and standard FSCIL settings across multiple benchmarks. The reported results and component ablations provide useful evidence of the method’s effectiveness.

Critical weaknesses
  1. Several central components lack the details needed to understand and reproduce the method. In Section 3.3, the descriptor q(Se,tce)q(S_e,t_{c_e}) is described conceptually, but its actual construction is not provided. Section 3.4 similarly does not explicitly define how the weighted covariance Σc\Sigma_c is computed. In addition, neither the value of const\mathrm{const} in Eq. (7) nor the number of retained eigenvectors kk in Eq. (8) is specified. These omissions make it difficult to understand the implementation and reproduce the reported results.

  2. The paper does not directly compare the learned distributional reference with a reference computed from the support samples, or spherical transfer with a simpler Euclidean transfer. The existing experiments examine recovery strategies, the number of recovered samples, and the VE components, but leave these two specific design choices less well understood. Comparing these alternatives while keeping the rest of the model unchanged would make their contributions clearer.

  3. Could the authors clarify the scope of class and domain novelty given CLIP’s broad pretraining? The novel classes and base-unseen domains are defined relative to the downstream sessions, but prior exposure during CLIP pretraining cannot be ruled out. The strong zero-shot results in Table 3 also indicate substantial pretrained knowledge. Please discuss how this affects the interpretation of learning novel classes from few examples, and whether any information is available about potential pretraining–evaluation overlap.

Questions for authors

Please see the questions raised in the weaknesses.

3 comments

Thank you for your thoughtful and encouraging feedback.

Response to W1: Implementation details

We thank the reviewer for highlighting these omissions. We agree that the original manuscript did not describe several implementation details sufficiently clearly. We provide the definitions and parameter values below and will incorporate them into the revised manuscript.

(1) Descriptor q(Se,tce)q(S_e,t_{c_e}).

We define the descriptor as

q(Se,tce)=[μ(Se);r(Se);tce],q(S_e,t_{c_e})=[\mu(S_e);r(S_e);t_{c_e}],

where semicolons denote concatenation. Here, μ(Se)\mu(S_e) summarizes the support distribution through its empirical kernel mean, r(Se)r(S_e) characterizes the pairwise visual similarity structure among support samples, and tcet_{c_e} provides the class text representation.

(2) Weighted covariance Σc\Sigma_c.

For each class cc, we map the observed support features and recovered features to the tangent space at the class text anchor tct_c, yielding ξi=Log⁡tc(zi)\xi_i=\operatorname{Log}_{t_c}(z_i). Each tangent feature receives the weight: We set wi=1w_i=1 for each observed support feature ziz_i. For each recovered feature zi=gjz_i=g_j, we set wi=ajw_i=a_j, where aja_j is obtained from Eq. (7). We compute the weighted mean and centered weighted covariance as

\frac{\sum_i w_i \xi_i}{\sum_i w_i}, \qquad \Sigma_c = \frac{\sum_i w_i(\xi_i-\bar{\xi}_c)(\xi_i-\bar{\xi}_c)^\top} {\sum_i w_i}.$$ The sums run over all observed support features and recovered features for class $c$. The leading eigenvectors of $\Sigma_c$ capture the dominant directions of class-specific variation in the tangent space. **(3) The constant in Eq. (7).** In the default setting, we retain five recovered features per novel class and set $\mathrm{const}=5$, so that their weights sum to five and their average weight is one. This constraint fixes their total weight, while individual weights are determined adaptively by their collective agreement with the predicted distributional reference. **(4) $k$ in Eq. (8).** We set $k=3$, retaining the eigenvectors corresponding to the three largest eigenvalues of $\Sigma_c$ as the principal class-specific variation directions used in Eq. (8). ## **Response to W2: Distributional reference and transfer geometry** We thank the reviewer for suggesting these controlled comparisons. We now report two ablations on DomainNet, changing only the component under investigation while keeping the remaining components and experimental settings unchanged. **(1) Learned distributional reference versus the support-based reference.** We replace the learned distributional reference $\hat{\mu}_c$ with the empirical kernel mean of the observed supports, $\mu(S_c)$, while retaining spherical transfer and all other components. As shown in Table R1, the learned reference improves average novel-class accuracy by **4.80 percentage points** and AHM by **2.39 percentage points**, with a 1.34-percentage-point decrease in average base-class accuracy. This comparison shows that predicting a distributional reference provides additional benefit over directly using the support-based reference for novel-class recognition and balanced base–novel performance. **Table R1. Learned versus support-based distributional references on DomainNet.** | Reference | Avg. Novel ↑ | Avg. Base ↑ | AHM ↑ | | :--- | :---: | :---: | :---: | | Support-based kernel mean | 61.27 | **83.42** | 70.49 | | Learned distributional reference (LeVa) | **66.07** | 82.08 | **72.88** | **(2) Spherical versus Euclidean transfer.** We replace spherical transfer with Euclidean transfer while retaining the learned distributional reference and all other components. As shown in Table R2, spherical transfer improves average novel-class accuracy, average base-class accuracy, and AHM by **5.22, 1.08, and 3.67 percentage points**, respectively. These results support the benefit of respecting the spherical geometry of normalized features when transferring historical intra-class variations. **Table R2. Spherical versus Euclidean variation transfer on DomainNet.** | Transfer geometry | Avg. Novel ↑ | Avg. Base ↑ | AHM ↑ | | :--- | :---: | :---: | :---: | | Euclidean | 60.85 | 81.00 | 69.21 | | Spherical (LeVa) | **66.07** | **82.08** | **72.88** |

Response to W3: Scope of novelty and the role of CLIP pretraining

We thank the reviewer for highlighting the need to clarify the scope of novelty.

(1) Novelty is defined relative to the downstream protocol. In our work, “novel classes” are classes introduced after the base session, and “base-unseen domains” are domains absent from downstream base training. Neither term implies absence from CLIP pretraining. We acknowledge that LeVa builds on pretrained visual–semantic knowledge, as reflected in the strong zero-shot results.

(2) Few-shot learning improves downstream recognition beyond the pretrained model. To clarify the remaining challenge, we report per-domain Top-1 accuracy at session 8 on DomainNet in Table R3. Zero-shot CLIP achieves 90.42% on Real but only 58.04% on Infograph and 19.05% on Quickdraw, showing that pretrained knowledge does not translate into equally effective recognition across domains. Using the same pretrained CLIP backbone, LeVa improves accuracy in every domain, including gains of 12.38 percentage points on Sketch and 22.35 percentage points on Quickdraw.

Table R3. Per-domain Top-1 accuracy (%) at session 8 on DomainNet. Both methods use the same pretrained CLIP backbone.

MethodRealClipartPaintingSketchInfographQuickdraw
Zero-shot CLIP90.4281.5376.9973.2758.0419.05
LeVa93.2386.2786.2785.6565.1741.40

Accordingly, “learning novel classes from few examples” should be interpreted as adapting a pretrained model to newly introduced downstream classes using limited labeled supports, while maintaining joint recognition of all seen classes—not as acquiring entirely unfamiliar concepts from scratch. The results above demonstrate the benefit of our full adaptation framework; the controlled comparisons with and without VR further isolate the contribution of variation recovery.

(3) Potential pretraining–evaluation overlap. The full pretraining dataset of the original CLIP model used in our experiments is not publicly available, so we cannot reliably audit overlap with our benchmarks at the image, class-concept, or domain level. The original CLIP paper analyzes sample overlap for its evaluated datasets [1], but does not establish the overlap status of our specific benchmarks and splits. Recent work further distinguishes sample overlap from domain coverage and shows that pretraining domain composition affects CLIP’s generalization performance [2].

We therefore cannot rule out pretraining–evaluation overlap. In particular, low zero-shot accuracy on a domain does not establish that it was absent from pretraining. Our comparisons using the same CLIP weights assess LeVa’s additional benefit given shared pretrained knowledge, rather than generalization to concepts or domains absent from pretraining. We will explicitly clarify this interpretation and limitation in the revised manuscript.

[1] Radford et al. Learning Transferable Visual Models From Natural Language Supervision. ICML, 2021.

[2] Mayilvahanan et al. In Search of Forgotten Domain Generalization. ICLR, 2025.

Dear Reviewer 1uJw,

Thank you again for your time and thoughtful review. As the discussion period closes on November 18 (AoE), we would be grateful to hear whether our responses have addressed your concerns. If any points remain unclear or require further discussion, please let us know, and we will respond promptly.

Best regards,

The Authors

Reviewer_Zhm1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a new problem setting, Open-Domain Few-Shot Class-Incremental Learning, which removes the assumption that base and incremental classes share a fixed domain configuration: novel-class few-shot samples may come from the base domain, unseen domains, or heterogeneous mixtures, with composition varying across classes and sessions. It proposes LeVa, a recover-and-exploit framework. Variation Recovery (VR) learns from base classes a predictor that maps sparse support evidence to a kernel-mean distributional reference, transfers historical intra-class deviations across class centers via parallel transport on the hypersphere, and selects compatible variations under semantic and distributional constraints. Variation Exploitation (VE) uses the recovered variations to build text-anchored directional variation models and combines visual, semantic, and directional scores. A new Hybrid-miniImageNet benchmark is constructed.

Critical strengths
  1. The OD-FSCIL formulation is a genuine extension and is more realistic. It is carefully distinguished from cross-domain FSCIL and hybrid incremental learning, and its "open-domain" means non-predefined composition, not globally unseen domains. This matches real deployment, where incoming classes arrive with unpredictable and often mixed domains, making a fixed source-to-target assumption the less realistic default.

  2. The recover-and-exploit framework is coherent. Treating base-class variations as prior knowledge to be selectively recovered, rather than directly transferred, is a sensible response to the risk that transferred deviations carry source-class content. The two-constraint selection (semantic + distributional) is well aligned with this concern.

Critical weaknesses
  1. The comparison set lacks the most natural ablation for the CLIP backbone. BiMC is the only CLIP-based baseline, and it is training-free with GPT-generated descriptions. A stronger comparison would include a CLIP-based method that also adapts on novel supports (e.g., a prototype-conditioned prompt baseline trained on the same 5-shot data without variation recovery). Without it, the gain over zero-shot CLIP in Table 3 may reflect adaptation in general rather than variation recovery specifically.

  2. The proposed method seems limited to specific backbones. More backbones should be tested to validate the generalizability of the method, such as ResNet or other variants of CLIP.

  3. The OD-FSCIL premise is defined but not empirically characterized. The paper asserts that domain composition is heterogeneous and non-predefined, but provides no statistics on the constructed benchmarks: how many sessions actually contain mixed-domain classes, what the per-class domain entropy is, and how often a single class's 5 shots span multiple domains. Without this, the reader cannot tell whether the benchmark instantiates the stated challenge or whether most novel classes are effectively single-domain, in which case OD-FSCIL collapses toward standard FSCIL.

  4. The gains may be driven by the benchmark's construction rather than by the claimed mechanism. LeVa's advantage on Hybrid-miniImageNet (+2.78 FA over BiMC) is smaller than on DomainNet (+9.14 FA). Since Hybrid-miniImageNet stylizes miniImageNet novel images, the base-domain visual statistics may remain unusually informative. The paper does not test a control where novel classes are drawn from genuinely disjoint domains, nor does it report performance as a function of the degree of domain heterogeneity per class. This makes it hard to attribute gains to variation recovery specifically.

  5. Standard FSCIL performance is shown only as curves, with no tabulated comparison. The paper evaluates on miniImageNet and CIFAR100 under the standard protocol (Section 4), but Fig. 3 reports only accuracy trends, and the main text gives a single final-accuracy number for miniImageNet (93.78% vs. 92.4% for BiMC) with no CIFAR100 numbers and no per-session table. This makes it impossible to determine whether LeVa preserves or sacrifices performance in the original single-domain setting.

Questions for authors

On Hybrid-miniImageNet, novel images are stylized versions of miniImageNet classes. How much of LeVa's gain survives if novel classes come from domains genuinely disjoint from the base domain? Please report performance as a function of per-class domain heterogeneity.

5 comments

Thank you for your thoughtful and encouraging feedback.

Response to W1: The comparison set lacks the most natural ablation for the CLIP backbone.

Thank you for suggesting this controlled comparison. The PCP result in the original Table 6 was obtained with VR enabled and therefore does not isolate the contribution of prompt adaptation alone. To address this concern, we now add a PCP-only baseline without VR and present two complementary comparisons. Both use the same CLIP backbone and the same 5-shot novel-class supports, with all other settings held fixed within each comparison.

(1) VR beyond prompt adaptation. As shown in Table R1, adding VR to the PCP-only baseline improves average novel-class accuracy, average base-class accuracy, and AHM by 0.83, 0.56, and 0.70 percentage points, respectively, with all other settings unchanged. This comparison isolates the additional benefit of VR beyond prompt adaptation.

(2) VR within the complete VE (Variation Exploitation) module? The recovered variations are used by the other VE components as well as PCP. We therefore also report the controlled comparison from Table 7, retaining all VE components—TAVM, VM, and PCP—and varying only whether recovered variations are used. As shown in Table R2, enabling VR improves average novel-class accuracy by 3.74 percentage points and AHM by 2.06 percentage points, while average base-class accuracy decreases by only 0.16 percentage points. This comparison isolates VR’s contribution within the full VE framework and shows a larger benefit when all VE components are available to exploit the recovered variations.

Table R1: Prompt adaptation with and without VR on Hybrid-miniImageNet.

MethodAvgNovel↑Avg_{Novel}\uparrowAvgBase↑Avg_{Base}\uparrowAHM↑\mathrm{AHM}\uparrow
PCP-only (without VR)82.4694.9888.17
PCP+VR83.2995.5488.87

Table R2: Complete VE with and without VR on Hybrid-miniImageNet.

MethodAvgNovel↑Avg_{Novel}\uparrowAvgBase↑Avg_{Base}\uparrowAHM↑\mathrm{AHM}\uparrow
Full VE without VR83.0496.5289.18
Full VE + VR (LeVa)86.7896.3691.24

Response to W2: The results of different backbones on RN-50 and VIT-B32 CLIP.

Thank you for suggesting evaluation with additional backbones. We have added experiments using CLIP RN50 and ViT-B/32, covering both CNN and Transformer visual architectures beyond the original ViT-B/16. For each backbone, we compare LeVa with BiMC† using the same backbone and evaluation protocol

As shown in Tables R3 and R4, LeVa consistently outperforms BiMC† across all eight incremental sessions on both datasets. With RN50, LeVa improves AHM by 3.79 and 10.46 percentage points on Hybrid-miniImageNet and DomainNet, respectively. With ViT-B/32, the corresponding improvements are 2.61 and 9.77 percentage points. These results show that LeVa’s effectiveness is not restricted to ViT-B/16 and extends to the additional CLIP backbones tested.

Table R3. Results with additional CLIP backbones on Hybrid-miniImageNet.

MethodBackboneBase(acc)S1(HM)S2(HM)S3(HM)S4(HM)S5(HM)S6(HM)S7(HM)S8(HM)AvgHM
BiMC†RN5090.5289.7486.4186.0582.2382.0780.2679.9579.6483.29
OursRN5093.8292.6589.7789.4886.2685.9784.1784.0884.2387.08
BiMC†ViT-B/3292.2792.0889.9089.6986.1085.8784.1283.8384.2286.98
OursViT-B/3295.4793.3592.2292.1589.2488.9386.9686.8687.0589.59

Table R4. Results with additional CLIP backbones on Domainnet.

MethodBackboneBase(acc)S1(HM)S2(HM)S3(HM)S4(HM)S5(HM)S6(HM)S7(HM)S8(HM)AvgHM
BiMC†RN5060.7057.6456.2355.3854.3253.3353.2452.4652.4854.39
OursRN5078.2769.0267.3365.7764.5363.6263.3962.5462.6464.85
BiMC†ViT-B/3266.4163.7562.8961.7860.7459.7259.6558.9559.0160.81
OursViT-B/3280.7175.5873.5971.7769.8968.7768.7468.1268.1770.58

Response to W3: Domain composition statistics of OD-FSCIL.

We agree that the original manuscript did not provide sufficient statistics to characterize how the OD-FSCIL setting is instantiated in our benchmarks. We thank the reviewer for highlighting this omission. We now supplement the experimental description with domain composition statistics for the actual 5-shot support sets used in our experiments. Each class’s five supports are sampled from its multi-domain training pool, without prescribing domain combinations or proportions and without explicitly enforcing domain mixing.

As shown in Table R5, 100% of the analyzed novel-class support sets span at least two domains: all 40 in Hybrid-miniImageNet and all 120 across the three DomainNet splits. Consequently, all eight incremental sessions in every split contain mixed-domain classes; in these splits, every novel class has mixed-domain supports.

We measure per-class support-set domain entropy as Hc=−∑dpc,dlog⁡2pc,dH_c = -\sum_d p_{c,d}\log_2 p_{c,d}, where pc,dp_{c,d} is the fraction of class cc’s five supports belonging to domain dd. The mean entropies are 1.48 ± 0.298 bits on Hybrid-miniImageNet and 1.56 ± 0.396 bits on DomainNet. Both the mean and standard deviation are computed across novel-class support sets and expressed in bits. These statistics confirm that within-class domain heterogeneity is present in the actual few-shot observations, rather than only in the underlying dataset.

Table R5. Domain composition of novel-class 5-shot support sets. Counts indicate the number of support sets spanning the specified number of distinct domains. Entropy is reported as mean ± standard deviation across novel-class support sets.

DatasetNovel-class support sets1 domain2 domains3 domains4 domains5 domainsDomain entropy (bits, mean ± SD)
Hybrid-miniImageNet400626801.48 ± 0.298
DomainNet (3 splits)120018554251.56 ± 0.396

Response to W4 and Questions: Evaluation with novel domains disjoint from the base domain and different domain compositions.

Thank you for raising this concern. The different gain magnitudes alone do not establish whether the improvement depends on benchmark construction: BiMC† reaches 87.99% final accuracy on Hybrid-miniImageNet but 64.24% on DomainNet, indicating substantially different baseline performance levels. To examine VR’s contribution more directly, we added a domain-disjoint control and compared LeVa with and without VR under different domain compositions.

Experimental setting. Base-class training and testing use only Real images, while novel-class support and test sets use exclusively non-Real domains. Within each incremental session, novel-class supports and test samples follow the same prescribed domain configuration.

For Hybrid-miniImageNet, excluding Real leaves three domains: Sketch, Painting, and Clipart. We construct seven incremental sessions covering all seven nonempty combinations of these domains. For DomainNet, we construct eight incremental sessions covering the five individual non-Real domains and three mixed-domain configurations:

Table R6. Novel-class domain configurations in the domain-disjoint control. Base classes use only the Real domain; novel-class support and test sets use the non-Real domain configuration specified for each session.

SessionHybrid-miniImageNetDomainNet
S1SketchSketch
S2PaintingPainting
S3ClipartClipart
S4Sketch + PaintingInfograph
S5Sketch + ClipartQuickdraw
S6Painting + ClipartSketch + Painting
S7Sketch + Painting + ClipartClipart + Infograph
S8—Quickdraw + Sketch

VR remains effective when novel domains are disjoint from the base domain. We compare full LeVa with LeVa w/o VR, which retains all VE components but does not use recovered variations. As shown below, VR improves AHM by 1.41 percentage points on Hybrid-miniImageNet and 3.15 percentage points on DomainNet, with positive gains in every incremental session. Thus, VR’s benefit persists when neither novel-class supports nor novel-class test samples contain Real-domain images. The DomainNet results additionally provide evidence beyond the construction of Hybrid-miniImageNet.

Table R7. Domain-disjoint evaluation on Hybrid-miniImageNet. Session-wise results are HM (%); AHM is averaged over S1–S7.

MethodS1S2S3S4S5S6S7AHM
ZS-CLIP90.3389.3589.2086.3086.1484.1183.2986.96
w/o VR97.5396.2196.1193.6893.7691.1790.8794.19
Ours97.9697.7496.9995.9194.8592.9992.7595.60

Table R8. Domain-disjoint evaluation on DomainNet. Session-wise results are HM (%); AHM is averaged over S1–S8.

MethodS1S2S3S4S5S6S7S8AHM
ZS-CLIP87.3985.7485.7881.4275.4176.7475.1972.8180.06
w/o VR91.9490.1289.7685.6383.8785.2083.9682.3486.60
Ours94.2792.6292.2689.1288.3288.7687.1585.5289.75

Performance under different domain compositions. We further group sessions by the number of domains in their incoming novel-class configuration. On Hybrid-miniImageNet, the mean session-wise HM gains from VR are 0.95, 1.71, and 1.88 percentage points for one-, two-, and three-domain configurations, respectively. On DomainNet, the corresponding gains are 3.05 and 3.31 percentage points for one- and two-domain configurations. These results show that VR benefits both single-domain and mixed-domain increments. The grouped comparisons describe performance under the evaluated configurations; they do not isolate domain heterogeneity from session order and class identity.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

We sincerely thank the Program Chairs, Senior Area Chairs, and Area Chairs for their efforts in organizing and coordinating the review process, and the reviewers for their careful assessment and constructive suggestions.

We appreciate the reviewers’ recognition of our contributions.

  • A realistic and well-motivated OD-FSCIL setting. All three reviewers (Zhm1, 1uJw, tp7F) acknowledged the motivation for studying few-shot class-incremental learning under heterogeneous domain composition.
  • A coherent recovery-and-exploitation framework. Reviewers Zhm1 and 1uJw recognized the rationale for recovering and exploiting historical intra-class variations, with Reviewer 1uJw emphasizing the complementary roles of the two modules. Reviewer tp7F acknowledged the novelty of spherical parallel transport.
  • Experimental evidence across multiple settings. Reviewers 1uJw and tp7F recognized the experimental evidence across open-domain and standard FSCIL settings.

We summarize our responses to the reviewers’ main concerns below.

  • Further validating VR’s contribution. Using the same CLIP backbone and 5-shot supports, adding VR improves performance with PCP alone and with all VE components. Additional experiments on DomainNet also show that distribution-guided selection outperforms random selection.
  • Expanding baseline and backbone comparisons. We evaluated more pretrained-model baselines, with LeVa achieving higher AHM than these baselines on both benchmarks. We also conducted additional experiments with CLIP RN50 and ViT-B/32.
  • Domain heterogeneity and domain-disjoint evaluation. We report support-set domain counts and entropy, showing that all analyzed novel-class support sets span multiple domains. Experiments with novel classes from domains absent during base training also show VR gains in sessions with single- and multi-domain novel classes.
  • Clarifying implementation and dataset construction. We provide additional computational details and parameter settings, together with dataset construction and test-set composition, to make the method and experimental setup clearer.

Detailed results and explanations are provided in our responses to individual reviewers. We will incorporate these additions into the revised manuscript and update the main text, table captions, and appendix accordingly.

We thank the reviewers for the suggestions that motivated these additional experiments and clarifications. We hope the new evidence helps address the main concerns and would appreciate its consideration in assessing our submission.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.