acceptodds
Under review as a conference paper at ICLR 2027

Superclass-guided Hierarchical Prompting for Few-Shot Class-Incremental Learning

Abstract

Few-Shot Class-Incremental Learning (FSCIL) aims at continuously learning new-class knowledge from extremely limited samples while preserving the recognition capability on old classes, which suffers from the dual challenges of catastrophic forgetting and overfitting. While existing prompt-based FSCIL methods achieve promising results, they overlook the semantic relationships among classes and fail to explicitly exploit the shared visual-semantic patterns across related classes. In this paper, we propose a Superclass-guided Hierarchical Prompt learning framework (SH-Prompt) to address this issue. Specifically, our method organizes base classes into semantic superclasses via a large language model (LLM) and progressively assigns newly introduced classes to the established superclasses. Based on the resulting semantic structure, we design a key-based superclass prompt pool to adaptively incorporate superclass-shared prompts for each sample through a semantic-driven retrieval mechanism. Furthermore, we derive a superclass-guided instance-level prompt learning loss, which introduces structured regularization across superclass semantics to improve the inter-class separability. Extensive experimental results on three FSCIL benchmark datasets demonstrate that our proposed method outperforms the state-of-the-art. Code will be released publicly.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual learning, Few-shot class-incremental learning

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_9qmKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SH-Prompt, which incorporates hierarchical semantic relationships into prompt-based Few-Shot Class-Incremental Learning (FSCIL). Building on SEC-Prompt, the method uses an LLM to partition classes into semantic superclasses. It injects superclass-shared prompts into deeper Transformer layers while keeping instance-level prompts in shallower layers, with both types optimized via superclass-guided loss terms (Lsp\mathcal{L}_{sp} and Lsip\mathcal{L}_{sip}). On CIFAR-100, ImageNet-R, and CUB-200, SH-Prompt shows modest improvements over SEC-Prompt and ASP, and includes robustness studies of partitioning strategies, superclass counts, and LLM backbones.

Critical strengths

Sharing superclass prompts across semantically related classes enables novel classes to reuse shared base-class representations, effectively easing the few-shot learning bottleneck. The key-similarity-based retrieval design is intuitive and practical, supporting the motivation.

Critical weaknesses
  1. Across Tables 1–3, the proposed method yields modest improvements over the strongest baseline, with average accuracy gains of only 0.5%, 1.4%, and 0.4% on CIFAR-100, ImageNet-R, and CUB-200, respectively. On CUB-200, harmonic accuracy (HAcc) improves by only 0.1%, while Performance Drop (PD) worsens from 3.3 to 3.8, highlighting an inconsistent trade-off across metrics. Furthermore, Table 5 shows that LLM-based grouping outperforms visual prototype-based grouping by a negligible 0.3% in average accuracy, with no gain in HAcc and worse PD. These results are insufficient to confirm whether these subtle gains are statistically meaningful or merely artifacts of variance.

  2. In Table 4, introducing superclass prompts simultaneously increases model capacity and adds deeper prompt tokens compared to the SEC-Prompt baseline. The current ablation fails to disentangle the benefit of semantic hierarchy from that of increased capacity. The authors should include a capacity-matched baseline—such as a random grouping strategy with an identical number and distribution of groups inserted at the same layer positions.

  3. There are noticeable inconsistencies both between the reported metrics and prior literature and internally within the manuscript. First, the baseline performance for SEC-Prompt diverges from the values reported in its original paper; the authors should clarify the exact reproduction setup or source for these numbers. Second, internal numerical discrepancies are present. For instance, in Table 1, the arithmetic mean of the per-session accuracies for the proposed method is ~89.71%, whereas the reported average is 89.5%. Additionally, for the full pipeline on ImageNet-R, Session 4 accuracy is stated as 79.1% in Tables 2 and 4 but drops to 78.1% in Tables 5, 6, and 7. The authors need to recheck and revise.

Questions for authors
  1. Discrepancies exist between the per-session accuracies and the reported average accuracies for LGSP and Ours in Table 1, and for Ours in Table 2. The authors should carefully recheck and revise these values. Furthermore, while the paper notes adherence to SEC-Prompt's protocols, could the authors clarify whether they directly cited the SEC-Prompt baseline numbers from the literature or reproduced them? If reproduced, were SEC-Prompt and the proposed method evaluated on identical few-shot sample splits, and what exact training hyperparameters were used?

  2. To isolate the genuine contribution of LLM-derived superclass semantics, could the authors provide a controlled comparison between semantic grouping and random grouping? Matching the random baseline in parameter count, number of clusters, and prompt insertion layers would clarify whether the observed gains stem from semantic relationships or from increased prompt capacity.

  3. Could the authors provide the exact superclass partitions and detailed implementation settings? For instance, 1) the explicit class-to-superclass mappings for each benchmark dataset and 2) the incremental assignment of novel classes to superclasses at each session.

  4. The definition and calculation of Performance Drop (PD) are ambiguous. While Appendix B defines PD as the accuracy degradation on base classes before and after incremental training, the values reported in the tables appear to reflect the overall accuracy delta between the first and final sessions. Could the authors clarify this formulation and how they derived the reported numbers?

3 comments

Thanks for your careful and valuable comments. We will explain your concerns point by point.
Q1: Across Tables 1–3, the proposed method yields modest improvements over the strongest baseline, with average accuracy gains of only 0.5%, 1.4%, and 0.4% on CIFAR-100, ImageNet-R, and CUB-200, respectively. On CUB-200, harmonic accuracy (HAcc) improves by only 0.1%, while Performance Drop (PD) worsens from 3.3 to 3.8, highlighting an inconsistent trade-off across metrics. Furthermore, Table 5 shows that LLM-based grouping outperforms visual prototype-based grouping by a negligible 0.3% in average accuracy, with no gain in HAcc and worse PD. These results are insufficient to confirm whether these subtle gains are statistically meaningful or merely artifacts of variance.
A1: We agree that the modest performance improvements reported in the original experiments warrant further statistical validation. To investigate whether these gains are consistent across different random initializations, we conduct additional experiments with five random seeds on all three benchmark datasets. We report the mean and standard deviation of each metric and paired t-test results. The results show that SH-Prompt consistently improves AavgA_{avg} over the strongest baseline across all three datasets, with gains of 0.56%, 0.44%, and 1.44% on CIFAR-100, CUB-200, and ImageNet-R, respectively. All three improvements are supported by paired t-tests (p < 0.01). HAcc also improves across all three datasets. Regarding PD, SH-Prompt achieves significant reductions on CIFAR-100 and ImageNet-R, whereas a slight increase is observed on CUB-200. As discussed in the paper, this may be related to the characteristics of CUB-200, where all categories belong to the bird class and already share substantial visual and semantic similarities. Consequently, superclass-level semantic grouping may provide relatively limited additional benefits for this fine-grained dataset, which may partly explain the lack of improvement in PD. Overall, these results provide statistical evidence that the improvements in Aavg are consistent across random seeds rather than merely arising from random variation. We will incorporate the corresponding results and statistical analysis into the revised paper. CIFAR-100 :

MetricSEC-PromptOursΔp-value(t)
AavgA_{avg}88.840±0.11489.400±0.1000.560.001705
HAcc83.480±0.13086.880±0.0843.4<0.001
PD8.340±0.1144.320±0.130-4.02<0.001

ImageNet-R :

MetricSEC-PromptOursΔp-value(t)
AavgA_{avg}76.220±0.11077.660±0.0551.44<0.001
HAcc71.300±0.18773.520±0.3032.22<0.001
PD10.500±0.0719.400±0.187-1.1<0.001

CUB200 :

MetricSEC-PromptOursΔp-value(t)
AavgA_{avg}84.520±0.08484.960±0.0890.440.001841
HAcc83.840±0.08984.140±0.1340.30.02305
PD3.360±0.1343.680±0.1640.320.040057

Note: p-values < 0.001 are reported as <0.001; all other p-values are reported with their original precision.

Q2: The authors should include a capacity-matched baseline—such as a random grouping strategy with an identical number and distribution of groups inserted at the same layer positions.
A2: To investigate whether the performance gain comes from the additional prompt capacity or the proposed hierarchical semantic modeling, we conduct an additional controlled experiment using randomly constructed superclass groups. Specifically, we keep the number of superclasses, prompt configuration, and training settings identical to those of SH-Prompt, while randomly assigning fine-grained classes to superclasses. As shown in the new results on ImageNet-R, the random grouping strategy achieves only marginal improvements over SEC-Prompt, increasing AavgA_{avg} and HAcc by 0.2% and 0.4%, respectively. In contrast, SH-Prompt achieves improvements of 1.2% in AavgA_{avg} and 1.6% in HAcc over random grouping, while also reducing PD by 0.8%. These results suggest that the performance gains cannot be attributed merely to the increased prompt capacity. Instead, organizing prompts according to meaningful superclass relationships plays an important role in improving the effectiveness of prompt learning. We will add this controlled experiment and the corresponding analysis to the revised paper.

method012345678910AavgA_{avg}HAccPD
SEC-Prompt82.880.379.777.376.575.574.173.773.673.172.376.371.310.5
random group82.980.479.877.276.57674.774.274.173.472.576.571.710.4
Ours83.381.080.778.578.177.476.275.875.574.973.777.773.39.6

Q3: There are noticeable inconsistencies both between the reported metrics and prior literature and internally within the manuscript. First, the baseline performance for SEC-Prompt diverges from the values reported in its original paper; the authors should clarify the exact reproduction setup or source for these numbers. Second, internal numerical discrepancies are present. For instance, in Table 1, the arithmetic mean of the per-session accuracies for the proposed method is ~89.71%, whereas the reported average is 89.5%. Additionally, for the full pipeline on ImageNet-R, Session 4 accuracy is stated as 79.1% in Tables 2 and 4 but drops to 78.1% in Tables 5, 6, and 7. The authors need to recheck and revise.
A3: We will carefully recheck the per-session accuracies and average accuracies in Tables, and correct the identified errors in the revised paper.
Q4: Discrepancies exist between the per-session accuracies and the reported average accuracies for LGSP and Ours in Table 1, and for Ours in Table 2. The authors should carefully recheck and revise these values. Furthermore, while the paper notes adherence to SEC-Prompt's protocols, could the authors clarify whether they directly cited the SEC-Prompt baseline numbers from the literature or reproduced them? If reproduced, were SEC-Prompt and the proposed method evaluated on identical few-shot sample splits, and what exact training hyperparameters were used?
A4: We will carefully recheck the per-session accuracies and average accuracies in the tables and correct the identified errors in the revised paper. Regarding the SEC-Prompt baseline, we reproduced it locally rather than directly citing the results reported in the original paper. Our reproduction followed the data settings and experimental protocols of the original SEC-Prompt paper. However, to ensure a fair comparison with ours, we set the random seed to 1 for both methods during reproduction. Apart from these adjustments, the remaining training configurations followed the original SEC-Prompt implementation. We will also clarify in the revised paper which baseline results are obtained through our own reproductions.
Q5: The definition and calculation of Performance Drop (PD) are ambiguous. While Appendix B defines PD as the accuracy degradation on base classes before and after incremental training, the values reported in the tables appear to reflect the overall accuracy delta between the first and final sessions. Could the authors clarify this formulation and how they derived the reported numbers?
A5: The PD values reported in our tables are calculated as the difference between the overall accuracy immediately after base training and that after all incremental sessions. We acknowledge that the definition in Appendix was inaccurate and will revise the corresponding formulation to clarify the calculation of PD.

Q6: Could the authors provide the exact superclass partitions and detailed implementation settings? For instance, 1) the explicit class-to-superclass mappings for each benchmark dataset and 2) the incremental assignment of novel classes to superclasses at each session.
A6: We provide the complete class-to-superclass mappings for all benchmark datasets. These details will be included in the appendix.

# Base-class-to-superclass mappings
CIFAR-100 : 
[0, 53, 57],
[1, 30, 32, 55],
[26, 45],
[3, 15, 19, 21, 31, 42, 43],
[4, 34, 36, 38, 50],
[27, 29, 44],
[6, 7, 14, 18, 24],
[2, 11, 35, 46],
[5, 20, 25],
[9, 10, 16, 28],
[22, 39, 40],
[8, 13, 41, 48, 58],
[12, 17, 37],
[23, 33, 49],
[47, 51, 52, 54, 56, 59],
ImageNet-R :
[0, 1, 2, 3],
[35, 36, 37],
[23, 24, 25, 31, 32, 33, 34],
[4, 5, 19],
[6, 7, 20, 21, 22],
[8, 9],
[10, 11, 12],
[13, 14, 15],
[16, 17, 18],
[27, 28],
[29, 30, 89],
[80, 81, 82, 83, 84, 85, 86, 87, 88],
[38, 39, 47, 48, 49, 50, 59, 63, 64, 67],
[40, 41, 42, 43, 44, 45],
[51, 52, 53, 68],
[46, 54, 55, 56, 57, 58, 60, 61, 62, 65, 66],
[72, 73, 74, 75, 76, 77],
[69, 70, 71, 78, 79],
[26, 90, 91, 92, 93, 94],
[95, 96, 97, 98, 99]
CUB200 : 
[0, 1, 2, 4, 5, 6, 7, 22, 23, 24, 43, 44, 57, 99],
[58, 59, 60, 61, 62, 63, 64, 65, 70, 71, 83],
[45, 49, 50, 51, 52, 85, 86, 88, 89],
[66, 67, 68, 69],
[78, 79, 80, 81, 82],
[35],
[36, 37, 38, 39, 40, 41, 42, 76, 77, 84],
[28, 29, 72, 73, 74, 92],
[27, 93],
[17, 18, 90],
[13, 14, 15, 16, 33, 34, 46, 47, 53, 54, 55, 56],
[20, 75],
[19],
[98],
[8, 9, 10, 11, 12, 25, 26, 48, 87, 94, 95, 96, 97],
[3, 21, 30, 31, 32, 91]
# Complete class-to-superclass mappings
CIFAR-100 : 
[0, 53, 57, 83],
[1, 30, 32, 55, 67, 72, 73, 91, 95],
[26, 45, 77],
[3, 15, 19, 21, 31, 42, 43, 88, 97],
[4, 34, 36, 38, 50, 63, 64, 65, 66, 74, 75, 80],
[27, 29, 44, 78, 93],
[6, 7, 14, 18, 24, 79, 99],
[2, 11, 35, 46, 98],
[5, 20, 25, 84, 94],
[9, 10, 16, 28, 61],
[22, 39, 40, 86, 87],
[8, 13, 41, 48, 58, 69, 81, 85, 89, 90],
[12, 17, 37, 68, 76],
[23, 33, 49, 60, 71],
[47, 51, 52, 54, 56, 59, 62, 70, 82, 92, 96]
ImageNet-R : 
[0, 1, 2, 3, 109, 110, 111],
[35, 36, 37, 120, 130, 155, 161, 168, 197, 198],
[23, 24, 25, 31, 32, 33, 34],
[4, 5, 19],
[6, 7, 20, 21, 22, 112, 115, 123, 124, 125, 128, 133, 134, 135, 137, 142, 143, 146, 147, 149, 150, 152, 158, 159, 167, 172, 174],
[8, 9, 114, 129, 138, 140, 141, 144, 151, 153, 156, 162, 175],
[10, 11, 12, 182, 183, 186],
[13, 14, 15],
[16, 17, 18, 122, 165],
[27, 28, 119, 121, 126, 127, 173, 176, 177, 178, 179, 180, 181, 184, 185, 187, 188, 189, 190, 191, 192, 193, 194, 195],
[29, 30, 89, 118, 157, 163, 170],
[80, 81, 82, 83, 84, 85, 86, 87, 88],
[38, 39, 47, 48, 49, 50, 59, 63, 64, 67],
[40, 41, 42, 43, 44, 45],
[51, 52, 53, 68],
[46, 54, 55, 56, 57, 58, 60, 61, 62, 65, 66],
[72, 73, 74, 75, 76, 77],
[69, 70, 71, 78, 79, 101, 102, 108],
[26, 90, 91, 92, 93, 94, 199],
[95, 96, 97, 98, 99, 100, 103, 104, 105, 106, 107, 113, 116, 117, 131, 132, 136, 139, 145, 148, 154, 160, 164, 166, 169, 171, 196]
CUB200 : 
[0, 1, 2, 4, 5, 6, 7, 22, 23, 24, 43, 44, 57, 99, 100, 105],
[58, 59, 60, 61, 62, 63, 64, 65, 70, 71, 83, 140, 141, 142, 143, 144, 145, 146],
[45, 49, 50, 51, 52, 85, 86, 88, 89],
[66, 67, 68, 69],
[78, 79, 80, 81, 82],
[35, 186, 187, 188, 189, 190, 191],
[36, 37, 38, 39, 40, 41, 42, 76, 77, 84, 101, 102, 103, 134, 135, 136, 137],
[28, 29, 72, 73, 74, 92, 106, 107, 110, 111],
[27, 93, 192, 193, 194, 195, 196, 197, 198],
[17, 18, 90, 148, 149, 184, 185],
[13, 14, 15, 16, 33, 34, 46, 47, 53, 54, 55, 56, 138, 139],
[20, 75, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 147],
[19, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 199],
[98, 108, 150, 151, 152, 153, 154, 155, 156],
[8, 9, 10, 11, 12, 25, 26, 48, 87, 94, 95, 96, 97, 133],
[3, 21, 30, 31, 32, 91, 104, 109]

We sincerely hope that our revisions and clarifications adequately address your concerns. We would be happy to provide further clarification if needed.

Reviewer_LQYqOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SH-Prompt, a superclass-guided hierarchical prompting framework for few-shot class-incremental learning (FSCIL). It leverages an LLM to organize classes into semantic superclasses, enabling knowledge sharing and reducing forgetting and overfitting. The paper introduces a key-based superclass prompt pool and a superclass-guided instance-level learning objective to enhance shared representations and inter-superclass discriminability.

Critical strengths
  1. The paper is well written and easy to follow. The pipeline of the methods is clear.
  2. The incorporation of LLM in the continuous learning task is a good exploration.
Critical weaknesses
  1. The paper has several writing and formatting issues, including spelling errors and improperly formatted references, such as “encod e” in the line 127, the citation without the information of year in line 107, AmA^m in line 180 but AmA_m in line 181.
  2. The analysis of proposed methods lacks sufficient depth and specificity, and the practical applicability remains limited. 3.The experiments part of the paper are insufficient, it lacks the running time analysis and detailed ablation studies.
Questions for authors
  1. Since LLM inference may cause a lot of time, what is the running time of the paper in training and inference stages comparing to other methods?
  2. In Table 4, the superclass prompt module improves ImageNet-R accuracy from 76.3% to 77.3%, but the lack of a capacity-matched baseline makes it unclear whether the gain stems from semantic modeling or additional prompts. Moreover, Table 5 shows only a marginal improvement from prototype-guided to LLM-guided partitioning (77.4% vs. 77.7%), with identical harmonic accuracy (73.3%). The differing superclass counts further confound the comparison. Controlled experiments with random or shuffled superclass assignments, matched prompt capacity, and identical superclass counts are needed to substantiate the contribution of hierarchical semantic modeling. 3.What is the definition of PiP_i and PjP_j in eq. 10, how can we get them?
  3. Does SH-Prompt support prompt pool expansion during incremental learning, similar to CODA-Prompt [1]? If the prompt pools remain fixed, how are existing prompts updated when new classes arrive, and how is interference with previously learned knowledge mitigated? How does the fixed superclass prompt pool handle novel classes whose semantics are poorly aligned with the existing superclasses? [1] Smith, James Seale, et al. "Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning." 2023 IEEE/CVF conference on computer vision and pattern recognition (CVPR). IEEE, 2023.
4 comments

Thanks for your careful and valuable comments. We will explain your concerns point by point.
Q1: The paper has several writing and formatting issues, including spelling errors and improperly formatted references, such as "encod e" in the line 127, the citation without the information of year in line 107, AmA^m in line 180 but AmA_m in line 181.
A1: In addition to addressing the issues mentioned above, we will carefully re-examine the paper and correct the identified spelling, formatting, and reference errors.
Q2: The analysis of proposed methods lacks sufficient depth and specificity, and the practical applicability remains limited.
A2: The key motivation of SH-Prompt is to exploit superclass-level semantic relationships as an intermediate knowledge structure for few-shot class-incremental learning. Unlike conventional prompt-based methods that primarily organize prompts at the individual class level, SH-Prompt introduces superclass-specific prompts to capture shared knowledge among semantically related classes. Specifically, superclass keys enable the model to retrieve relevant prompts based on visual features, while the super-class semantic regularization loss and inter-superclass loss impose complementary structural constraints to facilitate hierarchical knowledge learning. These components jointly support knowledge sharing across related classes while preserving superclass-specific representations. Regarding practical applicability, SH-Prompt accommodates evolving class sets by assigning novel classes to existing superclasses, without reconstructing the prompt pool. It reuses the shared knowledge learned from base classes and updates only the corresponding superclass prompts, enabling adaptation to novel categories with limited training data while retaining prompts associated with other superclasses.
Q3: The experiments part of the paper are insufficient, it lacks the running time analysis and detailed ablation studies.
A3: We have expanded the experimental evaluation by adding controlled random-grouping experiments, computational cost analysis, and further ablation studies to provide a more comprehensive assessment of SH-Prompt.
Q4: Since LLM inference may cause a lot of time, what is the running time of the paper in training and inference stages comparing to other methods?
A4: We clarify that the LLM is invoked only once for each task to construct the superclass partition during training, and add comparisons of training time, inference latency, and parameter counts between SEC-Prompt and SH-Prompt to quantify the additional computational overhead.
CIFAR-100 :

MethodTrainable Parameters(M)Total Parameters(M)Base Training Time / Epoch (s)Incremental Training Time / Epoch (s)Final Inference Time / Image (ms)
ASP2.288.0797.4-5.6
LGSP0.185.9348.541.04.7
SEC-prompt1.587.3171.10.52.7
Ours2.087.8280.80.74.7

ImageNet-R :

MethodTrainable Parameters(M)Total Parameters(M)Base Training Time / Epoch (s)Incremental Training Time / Epoch (s)Final Inference Time / Image (ms)
ASP5.090.8349.9-5.9
LGSP0.286.0159.126.55.3
SEC-prompt4.590.372.90.92.6
Ours5.291.0120.61.65.7

CUB200 :

MethodTrainable Parameters(M)Total Parameters(M)Base Training Time / Epoch (s)Incremental Training Time / Epoch (s)Final Inference Time / Image (ms)
ASP2.288.0129.1-5.5
LGSP0.286.060.911.04.8
SEC-prompt3.289.026.01.02.5
Ours3.889.633.41.54.9

Q5: In Table 4, the superclass prompt module improves ImageNet-R accuracy from 76.3% to 77.3%, but the lack of a capacity-matched baseline makes it unclear whether the gain stems from semantic modeling or additional prompts. Moreover, Table 5 shows only a marginal improvement from prototype-guided to LLM-guided partitioning (77.4% vs. 77.7%), with identical harmonic accuracy (73.3%). The differing superclass counts further confound the comparison. Controlled experiments with random or shuffled superclass assignments, matched prompt capacity, and identical superclass counts are needed to substantiate the contribution of hierarchical semantic modeling.
A5: To investigate whether the performance gain comes from the additional prompt capacity or the proposed hierarchical semantic modeling, we conduct an additional controlled experiment using randomly constructed superclass groups. Specifically, we keep the number of superclasses, prompt configuration, and training settings identical to those of SH-Prompt, while randomly assigning fine-grained classes to superclasses. As shown in the new results on ImageNet-R, the random grouping strategy achieves only marginal improvements over SEC-Prompt, increasing AavgA_{avg} and HAcc by 0.2% and 0.4%, respectively. In contrast, SH-Prompt achieves improvements of 1.2% in AavgA_{avg} and 1.6% in HAcc over random grouping, while also reducing PD by 0.8%. These results suggest that the performance gains cannot be attributed merely to the increased prompt capacity. Instead, organizing prompts according to meaningful superclass relationships plays an important role in improving the effectiveness of prompt learning. We will add this controlled experiment and the corresponding analysis to the revised paper.

method012345678910AavgA_{avg}HAccPD
SEC-Prompt82.880.379.777.376.575.574.173.773.673.172.376.371.310.5
random group82.980.479.877.276.576.074.774.274.173.472.576.571.710.4
Ours83.381.080.778.578.177.476.275.875.574.973.777.773.39.6

Q6: What is the definition of PiP_i and PjP_j in eq. 10, how can we get them?
A6: In Eq. (10), PiP_i and PjP_j denote the instance-level prompt representations corresponding to samples i and j, respectively. Following the D-Prompt setting in SEC-Prompt, the underlying prompt pool is randomly initialized and optimized during training to capture sample-specific feature information. Based on them, we introduce an inter-superclass loss to constrain the prompt feature distributions according to the superclass hierarchy.
Q7: Does SH-Prompt support prompt pool expansion during incremental learning, similar to CODA-Prompt [1]? If the prompt pools remain fixed, how are existing prompts updated when new classes arrive, and how is interference with previously learned knowledge mitigated? How does the fixed superclass prompt pool handle novel classes whose semantics are poorly aligned with the existing superclasses?
A7: Unlike CODA-Prompt, SH-Prompt adopts a fixed superclass-level prompt pool during incremental learning. This design is motivated by the few-shot setting of FSCIL: since only a limited number of samples are available for each novel class, we aim to leverage the shared knowledge learned from data-rich base classes rather than introduce new prompts that must be learned from very few samples. When a new class arrives, only the prompts associated with its assigned superclass are updated, while prompts belonging to other superclasses remain unchanged, thereby reducing cross-superclass interference. To support this design choice, we conduct two additional experiments that compare the fixed-pool strategy with two prompt expansion strategies based on random and weighted initialization, respectively. The fixed-pool strategy achieves better performance in our experiments, supporting this design choice. We will revise the paper and add these details to Appendix.

method012345678910AavgA_{avg}HAccPD
SEC-Prompt82.880.379.777.376.575.574.173.773.673.172.376.371.310.5
Random Init. of New Superclasses83.380.580.277.777.576.876.075.475.374.473.577.373.09.8
Weighted Init. of New Superclasses83.380.780.578.277.777.376.175.475.374.873.577.573.29.8
Ours83.381.080.778.578.177.476.275.875.574.973.777.773.39.6

We sincerely hope that our revisions and clarifications adequately address your concerns. We would be happy to provide further clarification if needed.

Thank you for the additional experiments and detailed clarifications. However, I still have concerns regarding the necessity of LLM. In Table 6, LLM-guided partitioning with 19 superclasses achieves 77.4% Aavg which is the same as Prototype-guided in Table 5. And it increases the computational overhead with limited improvement. Moreover, could the authors provide more details on the design of ND-Prompt? Because the authors address parts of my concerns, I will raise my score to 2.

We greatly appreciate your updated score and thoughtful feedback. We would like to clarify the purpose of the two comparisons. In Table 5, we compare the two partitioning strategies under their respective optimal configurations. Specifically, Prototype-guided partitioning determines the number of superclasses by selecting the K that achieves the highest average silhouette score based on cosine distances between visual class prototypes, resulting in 19 superclasses. In contrast, LLM-guided partitioning achieves its best performance with 20 superclasses. This comparison evaluates the two strategies under their respective selected configurations. To further isolate the contribution of LLM-provided semantic knowledge from the effect of superclass granularity, we conduct an additional controlled experiment by setting the number of superclasses in Prototype-guided partitioning to 20, matching the LLM-guided setting. Under this controlled setting, Prototype-guided partitioning achieves 77.3% AavgA_{avg} and 72.8% HAcc, whereas LLM-guided partitioning achieves 77.7% AavgA_{avg} and 73.3% HAcc, respectively. These results suggest that LLM-provided semantic knowledge provides additional benefits beyond visual prototype clustering. As for computational overhead, the LLM is invoked only during training; no LLM calls are made at inference. This training-time overhead is therefore reasonable and justified by the consistent performance gains. Regarding ND-Prompt, following the design in SEC-Prompt, we employ reverse attention with ReLU to emphasize dissimilar query-key pairs, thereby modeling non-discriminative information. Specifically, we reverse the cosine similarity between the query and key and retain positive responses using ReLU, as formulated in Eq. (6).

方法012345678910AavgA_{avg}PDHACC
SEC-Prompt82.880.379.777.376.575.574.173.773.673.172.376.310.571.3
Prototype-guided(K=19)83.080.780.078.177.577.176.175.675.574.673.777.49.373.3
Prototype-guided(K=20)82.980.580.178.177.577.176.075.575.174.373.277.39.772.8
LLM-guided(K=20)83.381.080.778.578.177.476.275.875.574.973.777.79.673.3
Reviewer_1cgLOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This article addresses the problem of catastrophic forgetting and overfitting in FSCIL, pointing out that existing prompt methods ignore inter class semantic relationships and shared visual semantic patterns. The proposed model uses LLM to organize base classes into semantic superclass and gradually classify new classes; designs a key based superclass prompt pool and adaptively introduces superclass shared prompts for samples through semantic driven retrieval; proposes instance level prompt learning loss. Three benchmark experiments have shown that it is superior to existing SOTA. On the one hand, the author's attitude towards this article is not serious enough, and they have not been able to organize and compare the latest relevant work and research results. On the other hand, in terms of the experimental records in the latest paper, this article can still achieve SOTA performance.

Critical strengths
  1. SH-Prompt utilizes a large language model to organize base classes into semantic superclass, gradually incorporating new classes into existing superclass, explicitly modeling semantic relationships between classes, and compensating for the shortcomings of existing prompt methods that ignore shared visual semantic patterns, providing structured priors for incremental learning.
  2. The paper achieves better results than existing SOTA on three FSCIL benchmark datasets, indicating that this method has strong effectiveness and generalization under standard protocols.
  3. Using LLM to establish superclass and support progressive inclusion of new classes, the framework is easy to extend to subsequent incremental tasks.
Critical weaknesses
  1. SH-Prompt uses the superclass structure generated by LLM as the organizational skeleton of the entire framework, but LLM may output suboptimal or even incorrect superclass partitioning. Once there is a deviation in the assignment of base superclass, the performance of the entire model will be affected.
  2. The introduction to the relevant work is somewhat outdated, failing to effectively summarize the complete context and cutting-edge achievements of the field. Missing references: [1] Few-Shot Class-Incremental Continual Learning-Based Fault Diagnosis of Photovoltaic Arrays Using Prompt-Guided Knowledge Distillation and I-V Curves. IEEE Internet Things J. 13(13): 29395-29407 (2026) [2] Few-shot class-incremental learning based on prompt guidance and multimodal fusion. Knowl. Based Syst. 342: 115862 (2026) [3]Quantized Residuals to Continuous Prompts for Few-Shot Class Incremental Learning in Vision-Language Models. CVPR 2026: 3856-3865 [4]Semantic-Guided Global-Local Collaborative Prompt Learning for Few-Shot Class Incremental Learning. CVPR 2026: 5511-5520
Questions for authors
  1. Hierarchical classification systems can often improve the accuracy of coarse-grained classification, but they are often prone to causing recognition of fine-grained categories, especially when subclasses within superclass are highly confused visually. Will the model in this article encounter this issue?
  2. LLM superclass construction, prompt pool retrieval, and structured regularization collectively increase training and inference overhead, but the paper does not report the time or parameter increments relative to the baseline.
2 comments

Thanks for your careful and valuable comments. We will explain your concerns point by point.
Q1: SH-Prompt uses the superclass structure generated by LLM as the organizational skeleton of the entire framework, but LLM may output suboptimal or even incorrect superclass partitioning. Once there is a deviation in the assignment of base superclass, the performance of the entire model will be affected.
A1: We acknowledge that superclass partitioning may affect the performance of SH-Prompt, as the superclass structure guides prompt organization and selection. We have discussed this limitation in the appendix. Meanwhile, our existing experiments with different LLMs and varying numbers of superclasses show relatively small performance variations, suggesting that SH-Prompt is not overly sensitive to the specific LLM or superclass configuration, and that current LLMs can provide reasonably effective superclass partitions for our task. In future work, we will explore more reliable superclass construction and validation strategies to mitigate the risk of suboptimal or even incorrect superclass partitioning.
Q2: The introduction to the relevant work is somewhat outdated, failing to effectively summarize the complete context and cutting-edge achievements of the field. Missing references: [1] Few-Shot Class-Incremental Continual Learning-Based Fault Diagnosis of Photovoltaic Arrays Using Prompt-Guided Knowledge Distillation and I-V Curves. IEEE Internet Things J. 13(13): 29395-29407 (2026) [2] Few-shot class-incremental learning based on prompt guidance and multimodal fusion. Knowl. Based Syst. 342: 115862 (2026) [3]Quantized Residuals to Continuous Prompts for Few-Shot Class Incremental Learning in Vision-Language Models. CVPR 2026: 3856-3865 [4]Semantic-Guided Global-Local Collaborative Prompt Learning for Few-Shot Class Incremental Learning. CVPR 2026: 5511-5520
A2: Thank you for your suggestion. We will update the Related Work section by incorporating the suggested recent studies, including the latest advances in prompt-based and vision-language-model-based FSCIL methods, to provide a more comprehensive overview of the field.
Q3: Hierarchical classification systems can often improve the accuracy of coarse-grained classification, but they are often prone to causing recognition of fine-grained categories, especially when subclasses within superclass are highly confused visually. Will the model in this article encounter this issue?
A3: Unlike conventional hierarchical classification, SH-Prompt uses superclass information to organize and select prompts rather than restrict classification to a specific superclass. The final prediction is still performed over the entire fine-grained category space. To further examine its impact on fine-grained recognition, we evaluate the classification accuracy of SEC-Prompt and SH-Prompt on test samples from each superclass, while retaining the full fine-grained classification space for both methods. As shown in Table, the average accuracy across superclasses increases from 74.2% to 74.9%. These results indicate that superclass-guided prompt learning does not generally compromise fine-grained recognition. Although slight performance drops are observed in some superclasses, the overall results demonstrate that the proposed method can effectively exploit shared superclass knowledge while preserving fine-grained discrimination.

012345678910111213141516171819avg
SEC-prompt70.064.972.777.971.067.969.775.473.670.779.579.885.073.379.586.373.172.465.075.374.2
ours70.466.076.178.768.665.972.175.475.569.278.480.386.374.383.986.874.174.667.373.274.9

Q4: LLM superclass construction, prompt pool retrieval, and structured regularization collectively increase training and inference overhead, but the paper does not report the time or parameter increments relative to the baseline.
A4: We add comparisons of training time, inference latency, and parameter counts between the baseline and SH-Prompt to quantify the additional computational overhead.
CIFAR-100 :

MethodTrainable Parameters(M)Total Parameters(M)Base Training Time / Epoch (s)Incremental Training Time / Epoch (s)Final Inference Time / Image (ms)
ASP2.288.0797.4-5.6
LGSP0.185.9348.541.04.7
SEC-prompt1.587.3171.10.52.7
Ours2.087.8280.80.74.7

ImageNet-R :

MethodTrainable Parameters(M)Total Parameters(M)Base Training Time / Epoch (s)Incremental Training Time / Epoch (s)Final Inference Time / Image (ms)
ASP5.090.8349.9-5.9
LGSP0.286.0159.126.55.3
SEC-prompt4.590.372.90.92.6
Ours5.291.0120.61.65.7

CUB200 :

MethodTrainable Parameter(M)Total Parameters(M)Base Training Time / Epoch (s)Incremental Training Time / Epoch (s)Final Inference Time / Image (ms)
ASP2.288.0129.1-5.5
LGSP0.286.060.911.04.8
SEC-prompt3.289.026.01.02.5
Ours3.889.633.41.54.9

We sincerely hope that our revisions and clarifications adequately address your concerns. We would be happy to provide further clarification if needed.

Comments on the paper

1 comment

We sincerely thank all reviewers for their valuable comments and constructive suggestions, which have helped us substantially improve the paper. We have uploaded the revised paper. In the revision, we have updated the Related Work section by incorporating the suggested recent studies, including the latest advances in prompt-based and vision-language-model-based FSCIL methods, so as to provide a more comprehensive overview of the field. We have also carefully re-examined the paper and corrected the identified spelling, formatting, and reference errors. In addition, we have carefully rechecked the per-session accuracies and average accuracies in the tables and corrected the identified errors; beyond average accuracies, we have also checked and revised other reported evaluation metrics to ensure their accuracy and consistency in the revised paper. We believe these revisions have improved the clarity, completeness, and rigor of the paper. We hope these revisions adequately address the raised concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.