Pith. sign in

REVIEW 3 major objections 5 minor 48 references

RainbowPrompt: Diversity-Enhanced Prompt-Evolving for Continual Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proposes a prompt-evolving mechanism that transforms and aligns task-specific prompts before merging them, and reports consistent accuracy gains over prior prompt-based continual learning on image and video benchmarks.

desk verdict RainbowPrompt reports consistent gains in prompt-based continual learning, but the class-incremental inference protocol is unspecified, making the headline results unverifiable until the test-time task-selection rule is described. read the letter →

arxiv 2507.22553 v1 pith:DUFBE66L submitted 2025-07-30 cs.CV cs.AIcs.LG

Kiseong Hong , Gyeong-hyeon Kim , Eunwoo Kim This is my paper
classification cs.CVcs.AIcs.LG
keywords continuallearningclass-incrementalprompt-basedpromptevolvingrepresentationdiversityvisiontransformervideoactionrecognitionlearnableprobabilisticgate
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Prompt-based continual learning keeps a frozen pre-trained model and tunes small prompts, but existing methods combine prompts in ways that narrow the representation they produce. This paper tries to establish that the missing ingredient is an evolution step: before merging all task-specific prompts into one unified prompt, transform and align them so that old knowledge is re-expressed in terms of the new task. The resulting RainbowPrompt is averaged from evolved base prompts, and a learned per-layer gate decides where to insert it. On image and video class-incremental benchmarks the paper reports average accuracy gains of 9.07% and 7.40% over prior methods, with lower forgetting, suggesting that diversity of the integrated prompt is what drives rehearsal-free continual-learning performance.

What carries the argument

The RainbowPrompt is the unified prompt at each layer, formed by averaging the evolved representations of all base prompts. Evolution is carried by attention-based transformation, with an inter-task affinity matrix $G=\mathrm{softmax}(QK^\top/\sqrt{d_k})$ for task-level reweighting and an inter-feature affinity matrix $F=\mathrm{softmax}(Q^\top K/\sqrt{d_k})$ for feature-level reweighting, and by task-guided alignment, a two-layer ReLU network that refines the transformed representations before averaging. A learnable probabilistic gate, a Bernoulli variable relaxed via Gumbel-Softmax, decides per task and per layer whether the RainbowPrompt is inserted. This machinery is what lets the method avoid both fixed-prompt weighting and generation from an entangled task-shared space.

What would settle it

Run the method on a class-incremental split in which test inputs arrive one at a time with no task label and no oracle access, and compare accuracy when the stored RainbowPrompt is chosen by the learned query function against accuracy when the true task identity is supplied; if the gap is large, the reported class-incremental gains rest on the unstated selection step rather than on the evolution mechanism.

Watch

Extended reading notes

Core claim

The central claim is that accumulated task-specific prompts should be actively evolved rather than simply selected or averaged. At each new task, RainbowPrompt appends a new base prompt, keeps earlier base prompts frozen, conditions them with a learnable task embedding, and applies two attention passes: one computing an inter-task affinity matrix over prompts and one computing an inter-feature affinity matrix over their dimensions, followed by a two-layer ReLU task-guided alignment. Averaging the aligned representations yields the RainbowPrompt for each layer, and a Bernoulli gate relaxed with Gumbel-Softmax selects which layers receive it. The paper argues this produces higher representational diversity, measured by the nuclear norm of the prompt, and shows across CIFAR-100, ImageNet-R, CUB, UCF-101, and ActivityNet that higher diversity tracks higher accuracy and lower forgetting.

Load-bearing premise

The method assumes that when a test input arrives and no task identity is given, the correct stored per-task RainbowPrompt can still be picked for that input, yet the paper does not describe or evaluate that selection step.

Editorial extensions

If this is right

  • The paper reports that RainbowPrompt beats the strongest prior method on every image and video benchmark it tests, with the largest margins on longer task sequences.
  • Its CIFAR-100 accuracy rises slightly from the 10-task to the 20-task setting while every baseline drops, so the benefit of prompt evolution appears to grow as more tasks accumulate.
  • The adaptive layer gate accounts for a measurable share of the gain: about 1.98% accuracy on ImageNet-R and 3.61% on CUBS over manual prompt placement.
  • Because 76.5% of the trainable parameters can be discarded at inference, the evolution machinery is a training-time cost rather than an inference-time overhead.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves the test-time selection of stored prompts unspecified; testing the query-based selection separately from oracle selection would show whether the method is genuinely class-incremental as declared.
  • The reported correlation between prompt nuclear norm and accuracy suggests a practical monitoring signal: track the evolving prompt's nuclear norm during training as an early warning that new-task knowledge is being diluted.
  • The evolution idea transfers to any setting where multiple parameter-efficient modules are learned sequentially and then merged, such as adapters or low-rank updates, so a natural extension is to test the transform-and-align step there.
  • Memory grows linearly with the number of stored task prompts; whether older prompts can be compressed once their knowledge is folded into evolved prompts is an open extension the paper does not address.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes RainbowPrompt, a rehearsal-free prompt-based continual learning method. For each new task, it keeps previously learned base prompts frozen, transforms the full accumulated prompt set through task-level and feature-level attention, aligns the transformed prompts with a small MLP, and averages them into a per-layer unified 'RainbowPrompt.' A learnable probabilistic gate selects which layers receive the prompt. The method is evaluated on image (ImageNet-R, CIFAR-100, CUBS) and video (UCF-101, ActivityNet) class-incremental benchmarks under 10-task and 20-task settings, reporting consistent average-accuracy and forgetting gains over L2P, DualPrompt, DualPrompt-PGP, CODA-Prompt, CPrompt, and ConvPrompt, together with ablations, sensitivity analyses, and an efficiency analysis.

Significance. If the evaluation protocol is valid, the empirical contribution is substantial: the method outperforms strong prompt-based baselines across eight benchmark configurations, including a 10.54% accuracy gain over ConvPrompt in the ActivityNet 20-task setting, and it does so while discarding 76.5% of its trainable parameters at inference. The paper also provides ablations for each proposed component, a sensitivity study, and an efficiency comparison. The strongest point is the breadth and consistency of the reported gains. However, the central claim cannot be accepted as stated because the test-time inference rule under the claimed class-incremental setting is not specified; the stored per-task RainbowPrompts and per-task gates have no described selection mechanism. The omission directly affects the validity of the headline comparison.

major comments (3)
  1. [Sec. 3.1, Sec. 3.2.2, Fig. 3, Eq. (7)] Section 3.1 states that the setting is class-incremental and that 'task identities are unknown during testing,' but Section 3.2.2 says the per-task RainbowPrompts are 'directly used for prediction' and the Fig. 3 caption says testing 'requires the RainbowPrompts for each task i ... appended to L_i selected layers determined by {g_i_l}.' The paper never specifies how a test sample of unknown task is assigned to one of the t stored RainbowPrompt/gate sets, or how multiple sets would be combined. The only candidate mechanism is the matching loss in Eq. (7), which trains a query function q(·) and task embeddings e_t, but no test-time retrieval or selection rule using q(·) is described. If the implementation used the true task ID, the experiments are task-incremental rather than class-incremental and the reported gains are not a valid comparison; if it used q(·), that procedure must be stated, defined, and evaluated. This is a load-bearing reproducibility gap.
  2. [Sec. 3.4, Sec. 4.2] The paper introduces a classifier φ in the optimization objective but does not specify whether it is a single unified classifier over all classes seen so far or a per-task classifier (one head per task). Class-incremental evaluation with unknown task identities requires the former, since the model must output over the union of all seen classes. If per-task heads are used, the setting again collapses to task-incremental learning. The paper should state the exact classifier architecture and how prediction logits are produced at test time, especially when a test sample is judged to belong to a particular task.
  3. [Fig. 2, Sec. 1] The diversity analysis that motivates the method is not well-defined. The base prompts are fixed parameters that do not depend on the input, yet the caption of Fig. 2a says the nuclear norm of the prompt is measured 'using test samples from all seen classes.' The paper should specify the exact quantity whose nuclear norm is computed (the prompt itself, the prompted attention output, or the resulting feature representation) and why the test input matters. Without this clarification, the central claim that 'higher representation diversity improves accuracy and reduces forgetting' is not falsifiable or reproducible.
minor comments (5)
  1. [Abstract, Sec. 5] The abstract reports 'average gains of 9.07% and 7.40%' whereas the conclusion reports an 'overall average margin of 8.23%.' Please clarify the aggregation: are these averages over all baselines or over the best baseline, and over which set of scenarios?
  2. [Table 3] The ablation table uses rows with '-' and checkmarks to indicate removed components, but the caption does not state that each row removes the marked component, making the first column header ambiguous.
  3. [Sec. 3.2.1, Eqs. (1)-(3)] The tensor shapes in the attention transformation are not fully specified: the dimensions of Q, K, V, the axes of P_l, and the resulting shapes of G, F, and the outputs should be stated explicitly so that the equations are unambiguous.
  4. [Sec. 3.2.1] The sentence 'we precede a task-conditioning step' should be reworded (e.g., 'we prepend a task-conditioning step'), and the broadcasting function σ(·) should be defined.
  5. [Sec. 3.2.2, Sec. 4.4] The paper says the prompt-evolving mechanism is performed only during training and that test-time predictions use stored RainbowPrompts, but the memory cost of storing one RainbowPrompt and one gate set per task is not discussed; a sentence on how this scales with the number of tasks would be useful.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation; central claim rests on independent benchmarks. Minor self-citation only; main weakness is an underspecified test-time selection rule, a correctness gap rather than circularity.

full rationale

The claimed gains are empirical comparisons on ImageNet-R, CIFAR-100, CUBS, UCF-101, and ActivityNet against external baselines. RainbowPrompt is constructed by Eqs. (1)-(5) from learnable projections, self-attention, task-conditioning, and an average; the diversity motivation (nuclear norm, Fig. 2) is measured independently of the training loss, since Eq. (7) contains no nuclear-norm term, so the diversity claim is not defined into the result. The only self-citation is [15], used for the generic idea that manual layer selection is impractical; the adaptive-prompting contribution is separately supported by the ablation in Tab. 3 and Fig. 5, so the citation is not load-bearing. No imported uniqueness theorem, no ansatz smuggled in via self-citation, and no fitted parameter relabeled as a prediction. The notable weakness is not circularity: Sec. 3.1 declares task identities unknown during testing, while Sec. 3.2.2 and the Fig. 3 caption state that stored per-task RainbowPrompts are used directly at test time, with no explicit retrieval rule. The matching loss q(·), e_t in Eq. (7) is trained, but no test-time selection procedure is specified. This is a reproducibility/correctness gap, not a derivation that reduces to its inputs.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central claim rests mostly on empirical training: it assumes frozen ViT prompt-tuning works, assumes the query-based matching loss can identify the right task knowledge at test without describing the procedure, and assumes that nuclear-norm diversity is a meaningful proxy for accuracy. The hand-set hyper-parameters (prompt length, Dp, Dn, loss weights, gate temperature, soft-decision epochs) are choices not derived from theory. No new physical or ontological entities are introduced.

free parameters (7)
  • prompt length L_p = 20
    Chosen as a single base prompt per task with length 20 across all scenarios; sensitivity analysis shows robustness, but the value itself is set by hand.
  • projection dimension D_p = 96 (10-task), 56 (20-task)
    Set differently per task setting in Sec. 4.1, suggesting tuning; it controls the capacity of the attention-based transformation.
  • weight matrix dimension D_n = 56 (10-task), 28 (20-task)
    Chosen per scenario for the task-guided alignment MLP; sensitivity shows low sensitivity but the value is hand-picked.
  • sparsity loss weight lambda_s = 0.01
    Set in Sec. 3.4 without ablations; it affects the behavior of the probabilistic gate.
  • matching loss weight lambda_m = 0.01
    Set in Sec. 3.4; it balances the task embedding matching regularizer.
  • Gumbel-softmax temperature tau = not reported
    Controls relaxation of the probabilistic gate; value not stated in the paper.
  • soft-decision training epochs = not reported
    The paper says soft decisions are used for 'a certain number of epochs' then hard sampling, but does not specify how many.
assumptions (4)
  • domain assumption Pre-trained ViT-B/16 remains frozen and prompt tuning can encode task knowledge
    The entire method inherits the standard prompt-based continual learning assumption; no analysis of when this fails is given.
  • domain assumption Query function q(·) from DualPrompt can provide a matching signal for task embeddings at training and at inference
    L_match uses q(x) and e_t; the paper does not detail test-time retrieval, so this assumption is load-bearing.
  • domain assumption Representational diversity measured by nuclear norm correlates with accuracy and low forgetting
    Motivated in Fig. 2, but the correlation is empirical and is not used as a training objective.
  • ad hoc to paper Two-level attention and task-guided alignment preserve previously learned knowledge without explicit replay
    No formal proof or theoretical analysis is given; the claim is supported only by ablations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RainbowPrompt: Diversity-Enhanced Prompt-Evolving for Continual Learning." pith.science (2026). https://pith.science/paper/DUFBE66L

@misc{pith2026250722553,
  author       = {Pith},
  title        = {Pith review of: RainbowPrompt: Diversity-Enhanced Prompt-Evolving for Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DUFBE66L}},
  note         = {Machine review of arXiv:2507.22553}
}
read the original abstract

Prompt-based continual learning provides a rehearsal-free solution by tuning small sets of parameters while keeping pre-trained models frozen. To meet the complex demands of sequential tasks, it is crucial to integrate task-specific knowledge within prompts effectively. However, existing works rely on either fixed learned prompts (i.e., prompts whose representations remain unchanged during new task learning) or on prompts generated from an entangled task-shared space, limiting the representational diversity of the integrated prompt. To address this issue, we propose a novel prompt-evolving mechanism to adaptively aggregate base prompts (i.e., task-specific prompts) into a unified prompt while ensuring diversity. By transforming and aligning base prompts, both previously learned and newly introduced, our approach continuously evolves accumulated knowledge to facilitate learning new tasks. We further introduce a learnable probabilistic gate that adaptively determines which layers to activate during the evolution process. We validate our method on image classification and video action recognition tasks in class-incremental learning, achieving average gains of 9.07% and 7.40% over existing methods across all scenarios.

Figures

Figures reproduced from arXiv: 2507.22553 by the authors.

Figure 1
Figure 1. Conceptual illustration of (a) existing prompt integra [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Results of the representation diversity in a class-incremental learning scenario with a 10-task setting on CIFAR-100. We compare [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the proposed framework. When a new task [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Results on CUBS. The average accuracy for all seen classes at each time step is shown with its standard deviation indicated by [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visualization of prompt insertion patterns across layers: manual prompting (red box) [ [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: , RainbowPrompt consistently outperforms the strong baselines, CODA-Prompt and ConvPrompt, across various [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 37 canonical work pages

  1. [1]

    Mem- ory aware synapses: Learning what (not) to forget

    Rahaf Aljundi, Francesca Babiloni, Mohamed Elho- seiny, Marcus Rohrbach, and Tinne Tuytelaars. Mem- ory aware synapses: Learning what (not) to forget. InProceedings of the European Conference on Com- puter Vision (ECCV), pages 139–154, 2018. 3

    work page 2018

  2. [2]

    Activitynet: A large-scale video benchmark for human activity un- derstanding

    Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity un- derstanding. InProceedings of the ieee conference on computer vision and pattern recognition, pages 961– 970, 2015. 2, 5

    work page 2015

  3. [3]

    Co2l: Contrastive continual learning

    Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. InProceedings of the IEEE/CVF International conference on computer vi- sion, pages 9516–9525, 2021. 3

    work page 2021

  4. [4]

    Efficient lifelong learning with a-gem.arXiv preprint arXiv:1812.00420, 2018

    Arslan Chaudhry, Marc’Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem.arXiv preprint arXiv:1812.00420, 2018. 3

    arXiv 2018

  5. [5]

    Morgan & Claypool Publishers, 2018

    Zhiyuan Chen and Bing Liu.Lifelong machine learn- ing. Morgan & Claypool Publishers, 2018. 3

    work page 2018

  6. [6]

    Towards discrim- inability and diversity: Batch nuclear-norm maximiza- tion under label insufficient situations

    Shuhao Cui, Shuhui Wang, Junbao Zhuo, Liang Li, Qingming Huang, and Qi Tian. Towards discrim- inability and diversity: Batch nuclear-norm maximiza- tion under label insufficient situations. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3941–3950, 2020. 2, 7

    work page 2020

  7. [7]

    A continual learn- ing survey: Defying forgetting in classification tasks

    Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ale ˇs Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learn- ing survey: Defying forgetting in classification tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3366–3385, 2021. 2

    work page 2021

  8. [8]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 1, 3, 6

    arXiv 2010

Show all 48 references
  1. [9]

    Catastrophic forgetting in connec- tionist networks.Trends in Cognitive Sciences, 3(4): 128–135, 1999

    Robert M French. Catastrophic forgetting in connec- tionist networks.Trends in Cognitive Sciences, 3(4): 128–135, 1999. 1, 2

    1999

  2. [10]

    Compact bilinear pooling

    Yang Gao, Oscar Beijbom, Ning Zhang, and Trevor Darrell. Compact bilinear pooling. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 317–326, 2016. 4

    2016

  3. [11]

    Consis- tent prompting for rehearsal-free continual learning

    Zhanxin Gao, Jun Cen, and Xiaobin Chang. Consis- tent prompting for rehearsal-free continual learning. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 28463– 28473, 2024. 2, 3, 6, 7, 8

    2024

  4. [12]

    Transformer feed-forward layers are key-value memories.arXiv preprint arXiv:2012.14913, 2020

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are key-value memories.arXiv preprint arXiv:2012.14913, 2020. 5

    2012 arXiv

  5. [13]

    Rethinking channel dimensions for efficient model design

    Dongyoon Han, Sangdoo Yun, Byeongho Heo, and YoungJoon Yoo. Rethinking channel dimensions for efficient model design. InProceedings of the IEEE/CVF conference on Computer Vision and Pat- tern Recognition, pages 732–741, 2021. 2

    2021

  6. [14]

    The many faces of robustness: A critical analysis of out- of-distribution generalization

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out- of-distribution generalization. InProceedings of the IEEE/CVF international c...

    2021

  7. [15]

    Exploration and exploitation in continual learning.Neural Networks, 188:107444, 2025

    Kiseong Hong, Hyundong Jin, Sungho Suh, and Eu- nwoo Kim. Exploration and exploitation in continual learning.Neural Networks, 188:107444, 2025. 2, 5

    2025

  8. [16]

    Re-evaluating continual learning scenarios: A categorization and case for strong baselines.arXiv preprint arXiv:1810.12488, 2018

    Y Hsu. Re-evaluating continual learning scenarios: A categorization and case for strong baselines.arXiv preprint arXiv:1810.12488, 2018. 3

    2018 arXiv

  9. [17]

    Compacting, picking and growing for unfor- getting continual learning.Advances in Neural Infor- mation Processing Systems, 32, 2019

    Ching-Yi Hung, Cheng-Hao Tu, Cheng-En Wu, Chien-Hung Chen, Yi-Ming Chan, and Chu-Song Chen. Compacting, picking and growing for unfor- getting continual learning.Advances in Neural Infor- mation Processing Systems, 32, 2019. 3

    2019

  10. [18]

    Categori- cal reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144, 2016

    Eric Jang, Shixiang Gu, and Ben Poole. Categori- cal reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144, 2016. 5

    2016 arXiv

  11. [19]

    Helpful or harmful: Inter-task association in continual learning

    Hyundong Jin and Eunwoo Kim. Helpful or harmful: Inter-task association in continual learning. InEuro- pean Conference on Computer Vision, pages 519–535. Springer, 2022. 3

    2022

  12. [20]

    Growing a brain with sparsity-inducing generation for continual learning

    Hyundong Jin, Gyeong-hyeon Kim, Chanho Ahn, and Eunwoo Kim. Growing a brain with sparsity-inducing generation for continual learning. InProceedings of the IEEE/CVF international conference on computer vision, pages 18961–18970, 2023. 1

    2023

  13. [21]

    Generating instance-level prompts for rehearsal-free continual learning

    Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 11847–11857, 2023. 3

    2023

  14. [22]

    Continual learning with node- importance based adaptive group sparse regulariza- tion.Advances in Neural Information Processing Sys- tems, 33:3647–3658, 2020

    Sangwon Jung, Hongjoon Ahn, Sungmin Cha, and Taesup Moon. Continual learning with node- importance based adaptive group sparse regulariza- tion.Advances in Neural Information Processing Sys- tems, 33:3647–3658, 2020. 1, 3

    2020

  15. [23]

    Self-regulating prompts: Foun- dational model adaptation without forgetting

    Muhammad Uzair Khattak, Syed Talal Wasim, Muza- mmal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foun- dational model adaptation without forgetting. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision, pages 15...

    2023

  16. [24]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 5

    2009

  17. [25]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 3

    2021 arXiv

  18. [26]

    Prefix-tuning: Op- timizing continuous prompts for generation.arXiv preprint arXiv:2101.00190, 2021

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Op- timizing continuous prompts for generation.arXiv preprint arXiv:2101.00190, 2021. 3

    2021 arXiv

  19. [27]

    Inflora: Interference- free low-rank adaptation for continual learning

    Yan-Shuo Liang and Wu-Jun Li. Inflora: Interference- free low-rank adaptation for continual learning. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 23638– 23647, 2024. 3

    2024

  20. [28]

    Pre-train, prompt, and predict: A systematic survey of prompt- ing methods in natural language processing.ACM Computing Surveys, 55(9):1–35, 2023

    Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompt- ing methods in natural language processing.ACM Computing Surveys, 55(9):1–35, 2023. 3

    2023

  21. [29]

    Gradi- ent episodic memory for continual learning.Advances in Neural Information Processing Systems, 30, 2017

    David Lopez-Paz and Marc’Aurelio Ranzato. Gradi- ent episodic memory for continual learning.Advances in Neural Information Processing Systems, 30, 2017. 3

    2017

  22. [30]

    Under- standing and mitigating overfitting in prompt tuning for vision-language models.IEEE Transactions on Circuits and Systems for Video Technology, 33(9): 4616–4629, 2023

    Chengcheng Ma, Yang Liu, Jiankang Deng, Lingxi Xie, Weiming Dong, and Changsheng Xu. Under- standing and mitigating overfitting in prompt tuning for vision-language models.IEEE Transactions on Circuits and Systems for Video Technology, 33(9): 4616–4629, 2023. 3

    2023

  23. [31]

    Locating and editing factual associations in gpt.Advances in Neural Information Processing Systems, 35:17359–17372, 2022

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt.Advances in Neural Information Processing Systems, 35:17359–17372, 2022. 5

    2022

  24. [32]

    Learning to re- member: A synaptic plasticity driven framework for continual learning

    Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to re- member: A synaptic plasticity driven framework for continual learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pages 11321–11329, 2019. 3

    2019

  25. [33]

    Prompt gradient projec- tion for continual learning

    Jingyang Qiao, Xin Tan, Chengwei Chen, Yanyun Qu, Yong Peng, Yuan Xie, et al. Prompt gradient projec- tion for continual learning. InThe Twelfth Interna- tional Conference on Learning Representations, 2023. 1, 2, 3, 6, 7

    2023

  26. [34]

    icarl: Incre- mental classifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incre- mental classifier and representation learning. InPro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017. 3

    2001

  27. [35]

    Convolutional prompting meets language models for continual learn- ing

    Anurag Roy, Riddhiman Moulick, Vinay K Verma, Saptarshi Ghosh, and Abir Das. Convolutional prompting meets language models for continual learn- ing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23616–23626, 2024. 1, 2, 3, 5, 6, 7, 8

    2024

  28. [36]

    Gradi- ent projection memory for continual learning.arXiv preprint arXiv:2103.09762, 2021

    Gobinda Saha, Isha Garg, and Kaushik Roy. Gradi- ent projection memory for continual learning.arXiv preprint arXiv:2103.09762, 2021. 1

    2021 arXiv

  29. [37]

    Continual learning with deep generative replay

    Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in Neural Information Processing Systems, 30, 2017. 3

    2017

  30. [38]

    Coda-prompt: Continual decomposed attention- based prompting for rehearsal-free continual learning

    James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention- based prompting for rehearsal-free continual learning. InProceedings of the IEE...

    2023

  31. [39]

    Ucf101: A dataset of 101 human ac- tions classes from videos in the wild.arXiv preprint arXiv:1212.0402, 2012

    K Soomro. Ucf101: A dataset of 101 human ac- tions classes from videos in the wild.arXiv preprint arXiv:1212.0402, 2012. 2, 5

    2012 arXiv

  32. [40]

    vclimb: A novel video class incremental learning benchmark

    Andr ´es Villa, Kumail Alhamoud, Victor Escor- cia, Fabian Caba, Juan Le ´on Alc ´azar, and Bernard Ghanem. vclimb: A novel video class incremental learning benchmark. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pages 19035–19044, 2022. 5

    2022

  33. [41]

    Pivot: Prompting for video continual learning

    Andr ´es Villa, Juan Le ´on Alc ´azar, Motasem Alfarra, Kumail Alhamoud, Julio Hurtado, Fabian Caba Heil- bron, Alvaro Soto, and Bernard Ghanem. Pivot: Prompting for video continual learning. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2023

  34. [42]

    The caltech-ucsd birds- 200-2011 dataset

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds- 200-2011 dataset. 2011. 2, 5

    2011

  35. [43]

    Tempo- ral segment networks for action recognition in videos

    Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaoou Tang, and Luc Van Gool. Tempo- ral segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740–2755, 2018. 6, 7

    2018

  36. [44]

    Dualprompt: Complementary prompting for rehearsal-free contin- ual learning

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free contin- ual learning. InEuropean Conference on Computer Vision, pages 631–648. Spring...

    2022

  37. [45]

    Learning to prompt for continual learning

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 139–149, 2...

    2022

  38. [46]

    Memory replay gans: Learning to generate new categories without for- getting.Advances in Neural Information Processing Systems, 31, 2018

    Chenshen Wu, Luis Herranz, Xialei Liu, Joost Van De Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without for- getting.Advances in Neural Information Processing Systems, 31, 2018. 3

    2018

  39. [47]

    Der: Dynamically expandable representation for class in- cremental learning

    Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class in- cremental learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pages 3014–3023, 2021. 3

    2021

  40. [48]

    Lifelong learning with dy- namically expandable networks.arXiv preprint arXiv:1708.01547, 2017

    Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dy- namically expandable networks.arXiv preprint arXiv:1708.01547, 2017. 3

    2017 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.