- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper proposes DualTrack, a replay-free framework for long-tailed class-incremental learning. It introduces a dual-stream architecture that separates Gaussian-based expert routing from LoRA adaptation. The framework further incorporates adaptive expert expansion and prototype-based knowledge transfer to improve the learning of tail classes while preserving historical knowledge. Experiments on six benchmarks show competitive performance. Ablation studies and robustness evaluations provide additional validation of the proposed method.
Critical strengths
-
The paper provides a relatively comprehensive experimental evaluation across six benchmarks, demonstrating competitive performance against multiple representative continual learning baselines. The ablation studies offer useful evidence for several major components.
-
The proposed route-aware prototype transfer is a reasonable approach to addressing unreliable feature estimates for tail classes. By incorporating visual similarity, expert-route overlap, and class-frequency-dependent weighting, the method enables knowledge transfer from historical classes without storing previous images.
Critical weaknesses
-
The paper's presentation and writing could be improved. Figure 1 contains excessive information and is difficult to follow, while Figure 3 uses a somewhat informal visual style and contains a spelling error "Guassian router". The ordering of baselines in Table 1 is also somewhat confusing and lacks a clear rationale, organizing them chronologically by publication year would make the comparison more intuitive. In addition, the Ethics Statement section is left empty, and some descriptions in the appendix are repetitive, suggesting that the manuscript would benefit from more careful proofreading and editing.
-
The discussion of related work is insufficient in two aspects. First, the subsection Continual learning and parameter-efficient adaptation only covers works published in or before 2024, overlooking recent advances. Although DualPrompt and CL-LoRA are discussed elsewhere, the distinctions between DualTrack and existing parameter-efficient adaptation strategies are not sufficiently clarified. Second, given the central role of prototype-based prediction and tail-class knowledge transfer in this paper, relevant prototype-based continual learning methods, such as LADA [1] DIA [2], and GPA [3], should also be discussed. A more thorough comparison with these approaches especial the prototype subcomponents would help better establish the novelty and technical contributions of DualTrack.
-
The evaluation of long-tailed learning remains limited in terms of imbalance severity and dataset diversity. All six main benchmarks use the same imbalance ratio of 10, which is relatively mild compared with more challenging long-tailed recognition settings. Experiments on more challenging or naturally imbalanced datasets, such as Places-LT or iNaturalist, under suitable class-incremental protocols would provide stronger evidence for the method's generalizability. More detailed head-, medium-, and tail-class improvements would also help clarify its effectiveness under severe imbalance.
[1] LADA: Scalable Label-Specific CLIP Adapter for Continual Learning. ICML 2025
[2] Dynamic Integration of Task-Specific Adapters for Class Incremental Learning. CVPR 2025
[3] A Tiny Change, A Giant Leap: Long-Tailed Class-Incremental Learning via Geometric Prototype Alignment. ICCV 2025
Questions for authors
-
Section 4 reports a prototype-fusion weight of 1.0, while Appendices B.5 and B.9 describe a neutral prototype-only inference mode in which Specific expert outputs are disabled. Could the authors clarify the exact inference configurations used for the main results? In particular, how does expert routing directly contribute to the final predictions when neutral prototype-only inference is adopted?
-
Appendices B.3 and B.9 indicate that the reference configuration uses old_lr_mult = 0, implying that eligible historical experts receive no effective parameter updates even when historical gradient projection is applied. Please clarify whether gradient-direction protection is actually active in the reported experiments and provide evidence for its contribution?
9 comments
Q2:Insufficient Related Work
We have added more recent models, including TUNA and CoRE. We could not make direct performance comparisons with methods such as LADA, DIA, and GPA because the code was not sufficiently clear.But we found more recent research on extending expert models in class-incremental scenarios[1, 2, 3]: CoRE, TUNA, and SEMA.
| Dataset | Long-tail setting | Method | Final accuracy (%) | Wall-clock time (min) | Peak process GPU memory (MiB) | Final trainable parameters (M) |
|---|---|---|---|---|---|---|
| CUB200 | CUB-LT | TUNA | 83.83 | 37.24 | 4,894 | 3.959 |
| CUB200 | CUB-LT | CaRE | 83.19 | 62.37 | 21,816 | 2.152 |
| CUB200 | CUB-LT | SEMA | 69.14 | 41.00 | 4,830 | 0.154 |
| CIFAR-100 | IR=10 | TUNA | 85.45 | 136.89 | 2,242 | 3.806 |
| CIFAR-100 | IR=10 | CaRE | 85.57 | 335.55 | 14,848 | 1.906 |
| CIFAR-100 | IR=10 | SEMA | 81.37 | 88.12 | 7,176 | 0.077 |
| CIFAR-100 | IR=100 | TUNA | 85.18 | 111.39 | 2,250 | 3.806 |
| CIFAR-100 | IR=100 | CaRE | 85.09 | 312.35 | 15,030 | 1.906 |
| CIFAR-100 | IR=100 | SEMA | 78.10 | 83.81 | 7,130 | 0.077 |
| TinyImageNet | IR=10 | TUNA | 87.51 | 87.05 | 4,192 | 3.959 |
| TinyImageNet | IR=10 | CaRE | 88.00 | 701.27 | 15,014 | 2.152 |
| TinyImageNet | IR=10 | SEMA | 84.38 | 432.52 | 7,828 | 0.154 |
| TinyImageNet | IR=100 | TUNA | 87.18 | 81.75 | 4,192 | 3.959 |
| TinyImageNet | IR=100 | CaRE | 87.53 | 587.36 | 15,090 | 2.152 |
| TinyImageNet | IR=100 | SEMA | 82.80 | 282.81 | 9,692 | 0.154 |
| CUB200 | CUB-LT | Dual | 86.27 | 16.44 | 3,440 | 0.369 |
| CIFAR-100 | IR=10 | Dual | 85.01 | 61.37 | 1,980 | 0.319 |
| CIFAR-100 | IR=100 | Dual | 84.79 | 52.33 | 1,980 | 0.319 |
| TinyImageNet | IR=10 | Dual | 88.27 | 57.23 | 3,280 | 0.369 |
| TinyImageNet | IR=100 | Dual | 87.73 | 50.76 | 3,280 | 0.369 |
Note: CUB-LT uses linear long-tailed sampling with a minimum retention ratio of 0.1. After discrete sampling, the maximum and minimum class counts are 30/2, yielding an actual ratio of 15:1.
Q3:Limited Evaluation
We have added a more challenging 1:100 imbalance setting and have also conducted experiments on Places-LT, a naturally imbalanced dataset. We report the results by head, middle, and tail groups:
| Dataset | Method | Seed | H 1--3 (per task; mean) | M 4--6 (per task; mean) | T 8--10 (per task; mean) | Mean across all 10 tasks |
|---|---|---|---|---|---|---|
| cifar100 | DualTrack | 1993 | 85.0 / 85.9 / 86.7; 85.80 | 86.3 / 84.9 / 74.6; 81.93 | 85.2 / 87.0 / 87.2; 86.48 | 84.73 |
| cifar100 | cl_lora | 1993 | 68.7 / 63.8 / 63.2; 65.23 | 76.8 / 85.0 / 69.8; 77.20 | 84.0 / 86.5 / 87.8; 86.10 | 76.03 |
| cifar100 | keeplora | 1993 | 91.4 / 75.3 / 72.1; 79.60 | 81.8 / 79.2 / 74.6; 78.53 | 81.1 / 73.9 / 79.0; 78.00 | 78.78 |
| cifar100 | CaRE | 1993 | 85.7 / 87.4 / 89.1; 87.40 | 86.6 / 84.8 / 72.4; 81.27 | 82.2 / 85.5 / 87.4; 85.03 | 85.09 |
| cifar100 | SEMA | 1993 | 83.6 / 75.7 / 88.7; 82.67 | 80.4 / 80.6 / 49.1; 70.03 | 74.5 / 84.8 / 84.4; 81.23 | 78.10 |
| cifar100 | TUNA | 1993 | 85.6 / 86.2 / 89.2; 87.00 | 86.7 / 84.0 / 74.1; 81.60 | 84.3 / 86.5 / 87.3; 86.03 | 85.18 |
| tinyimagenet | CaRE | 1993 | 87.6 / 88.9 / 85.8; 87.43 | 85.5 / 87.6 / 86.0; 86.37 | 86.7 / 90.5 / 85.9; 87.70 | 87.53 |
| tinyimagenet | SEMA | 1993 | 82.3 / 79.7 / 78.3; 80.10 | 75.0 / 84.5 / 78.4; 79.30 | 82.9 / 92.2 / 86.6; 87.23 | 82.80 |
| tinyimagenet | TUNA | 1993 | 88.0 / 88.7 / 85.0; 87.23 | 84.3 / 86.9 / 85.0; 85.40 | 86.1 / 91.3 / 86.1; 87.83 | 87.18 |
| tinyimagenet | DualTrack | 1993 | 88.6 / 90.2 / 86.7; 88.50 | 85.5 / 87.6 / 86.2; 86.40 | 88.2 / 90.7 / 85.6; 88.16 | 87.73 |
| Method | Full-set final task-macro (%) | Incremental task-macro mean (%) | Train time (min) | Peak allocated GPU memory (MiB) | Trainable parameters (M) |
|---|---|---|---|---|---|
| DualTrack | 46.9198 | 56.6976 | 264.51 | 2,449 | 0.447 |
| TUNA | 45.5002 | 56.4520 | 207.76 | 3,529 | 4.213 |
| CaRE | 46.0872 | 56.6488 | 986.98 | 9,856 | 2.553 |
| cl_lora | 35.1324 | 46.4400 | 283.18 | 8,763 | 0.276 |
| kp_lora | 38.7960 | 47.4512 | 210.44 | 4,851 | 0.983 |
Q4:Inference Configuration and Gradient Protection
The ablation results for the prototype fusion weight are reported below for CUB200 with seed=2026.
| Setting | Full-set micro (%) | Full-set class-macro (%) |
|---|---|---|
| inference, final features: Specific scale=0 | 36.2433 | 36.9232 |
| inference, final features: Specific scale=0.5 | 86.2260 | 86.9065 |
| inference, final features: Specific scale=1 | 77.8737 | 78.0313 |
| Training with Specific residual on | 86.2260 | 86.9065 |
| Training with Specific residual off | 31.2910 | 30.7290 |
| Prototype fusion weight | Full-set micro | Full-set class macro | Full-set task mean |
|---|---|---|---|
| 0.0 | 70.6489% | 70.8898% | 70.8056% |
| 0.5 | 76.3048% | 76.6063% | 76.5482% |
| 1.0 | 85.9860% | 86.3065% | 86.9103% |
We apologize for the confusion caused by the appendix and will clarify the inference procedure and the prototype fusion weight. Classification uses two paths. One computes cosine similarities to the stored class prototypes, and the other uses a linear classification head. The two outputs are added to obtain the final prediction vector, as follows:
We thank Reviewer QNJe for the valuable comments. In the revised manuscript, we have added recent related models to the main comparison table, included performance overhead experiments on additional datasets, provided more detailed ablation studies, and introduced experiments under more extreme imbalance scenarios. Furthermore, we have simplified the framework diagram in the main text, retaining only an intuitive illustration of the core mechanisms and data flow. Meanwhile, we have provided more detailed and explicit explanations of our model's mechanisms in the appendix to reduce potential misinterpretations. By demonstrating the rationality of our architectural design with experimental data, we have further elevated the rigor of the paper.
Q1:Presentation Issues
Regarding the presentation of the paper, Figure 1 clarifies the data flow so that readers can follow the entire training process. Figure 3 illustrates how routing decisions are made during training and inference and how they differ between the two stages. We have improved the color scheme. (Spelling errors.) The revised main table will list the models by year. We have also added an ethics statement and clarified the descriptions in the appendix.