arxiv: 2604.11091 · v1 · submitted 2026-04-13 · 💻 cs.CV

Recognition: no theorem link

LDEPrompt: Layer-importance guided Dual Expandable Prompt Pool for Pre-trained Model-based Class-Incremental Learning

Linjie Li , Zhenyu Wu , Huiyu Xiao , Yang Ji

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:44 UTC · model grok-4.3

classification 💻 cs.CV

keywords class-incremental learningprompt tuningpre-trained modelscontinual learningexpandable prompt poollayer importanceparameter-efficient learning

0 comments

The pith

LDEPrompt guides prompt selection by layer importance and lets the prompt pool expand or freeze dynamically to handle new classes in pre-trained model incremental learning.

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

Existing prompt-based class-incremental learning methods rely on fixed pools of prompts, require manual choices, and lean heavily on the frozen backbone for matching. The paper introduces LDEPrompt to replace these with an adaptive mechanism that scores layer importance to pick prompts and then freezes or grows the pool as tasks arrive. Experiments on standard benchmarks show this yields higher accuracy than prior approaches, indicating the method scales better when classes are added sequentially without retraining the entire model.

Core claim

The layer-importance guided dual expandable prompt pool performs adaptive layer selection together with dynamic freezing and expansion of the prompt pool, overcoming fixed-pool size, manual selection, and backbone dependence while delivering state-of-the-art accuracy on class-incremental learning benchmarks.

What carries the argument

The dual expandable prompt pool, which scores layer importance to choose and then freezes or adds key-prompt pairs on the fly.

If this is right

The prompt pool grows only when new classes require it, limiting unnecessary parameters.
Freezing earlier prompts protects performance on old tasks while new ones are learned.
Layer-specific importance reduces dependence on the entire pretrained backbone for every decision.
The approach scales to longer task sequences without manual redesign of pool size.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same layer-importance scoring could be applied to other continual-learning settings such as domain-incremental image classification.
If expansion cost remains low, the method might extend to very large foundation models where full fine-tuning is impractical.
Testing on video or multimodal data would check whether layer importance transfers across modalities.

Load-bearing premise

That measuring layer importance will consistently pick the right prompts and that controlled pool expansion will stay stable without raising compute cost or forgetting earlier classes.

What would settle it

Running LDEPrompt on a standard 10-task split of CIFAR-100 and finding its final accuracy lower than a fixed-pool baseline such as L2P or DualPrompt.

read the original abstract

Prompt-based class-incremental learning methods typically construct a prompt pool consisting of multiple trainable key-prompts and perform instance-level matching to select the most suitable prompt embeddings, which has shown promising results. However, existing approaches face several limitations, including fixed prompt pools, manual selection of prompt embeddings, and strong reliance on the pretrained backbone for prompt selection. To address these issues, we propose a \textbf{L}ayer-importance guided \textbf{D}ual \textbf{E}xpandable \textbf{P}rompt Pool (\textbf{LDEPrompt}), which enables adaptive layer selection as well as dynamic freezing and expansion of the prompt pool. Extensive experiments on widely used class-incremental learning benchmarks demonstrate that LDEPrompt achieves state-of-the-art performance, validating its effectiveness and scalability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LDEPrompt adds layer-importance guidance and dual expandable prompt pools to address fixed-size and manual-selection limits in prompt-based class-incremental learning, but the SOTA claim is hard to evaluate without controls on parameter growth.

read the letter

The main point is that this paper takes existing prompt-pool approaches for class-incremental learning and makes the pool dual and expandable, with selection guided by layer importance. It targets the problems of fixed pools, manual prompt choice, and heavy backbone dependence that the abstract lays out clearly. That combination is the concrete new element here, and the framing of those limitations is straightforward and useful for anyone working in this area. The paper does a decent job connecting the design choices back to those gaps without overclaiming novelty in the core prompt idea itself. On the soft spots, the abstract gives no numbers at all on final pool sizes, per-task expansion rates, total trainable parameters, or FLOPs compared to fixed-pool baselines. This leaves the stress-test concern standing: if expansion is only lightly constrained, the reported gains could trace to added capacity rather than the layer-guidance or dual-pool mechanics. Without ablations that hold capacity constant or show bounded growth, the central claim stays under-supported even if the full paper has the experiments. The work is aimed at researchers already using prompt tuning for vision continual learning. Someone in that group would pick up the specific architecture details and see how they try to scale the pool. It deserves peer review because the problem is real and the proposal is specific enough to test, though any referee should be asked to check the capacity controls and attribution of results first. I would send it out rather than desk-reject.

Referee Report

2 major / 2 minor

Summary. The paper proposes LDEPrompt, a prompt-based method for class-incremental learning (CIL) with pre-trained vision models. It introduces layer-importance guided adaptive layer selection together with a dual expandable prompt pool that performs dynamic freezing and expansion of prompts. The approach aims to overcome the limitations of fixed-size prompt pools, manual prompt selection, and heavy reliance on the backbone network for matching. The authors report extensive experiments on standard CIL benchmarks (e.g., CIFAR-100, ImageNet-100) claiming state-of-the-art accuracy while demonstrating scalability through the adaptive mechanisms.

Significance. If the performance gains are shown to arise specifically from the layer-importance guidance and controlled dual expansion rather than from net increases in trainable prompt capacity, the work would offer a practically useful advance in prompt-based continual learning. It extends prior prompt-pool methods by adding explicit adaptivity and could influence designs that seek to balance plasticity and stability without unbounded parameter growth. The empirical focus on widely used benchmarks makes the contribution directly testable.

major comments (2)

[§4.2, Tables 2-3] §4.2, Table 2 and Table 3: The SOTA claims are presented without reporting the final prompt pool size, total trainable prompt parameters, or per-task expansion counts for LDEPrompt versus the fixed-pool baselines (e.g., L2P, DualPrompt). This omission leaves open the possibility that accuracy improvements are driven by increased effective capacity from dual expansion rather than the layer-importance guidance or freezing mechanism, directly undermining attribution of the central claim.
[§3.3, §4.3] §3.3 and §4.3: The layer-importance scoring and dual-pool expansion rules are described, yet the ablation studies do not include a controlled comparison that holds total prompt capacity fixed while varying only the guidance and expansion logic. Without such an isolation experiment, it is impossible to confirm that the proposed mechanisms, rather than simply more prompts, produce the reported gains.

minor comments (2)

[Figure 2] Figure 2: The diagram of the dual expandable pool would benefit from explicit annotation of which components are frozen versus trainable at each stage to clarify the dynamic freezing process.
[§2] §2 Related Work: A few recent prompt-based CIL papers (post-2023) are missing from the discussion; adding them would better situate the novelty of the dual-expansion strategy.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. The comments highlight important aspects of transparency and attribution that we address below. We have revised the manuscript to include additional reporting and experiments that directly respond to the concerns raised.

read point-by-point responses

Referee: [§4.2, Tables 2-3] §4.2, Table 2 and Table 3: The SOTA claims are presented without reporting the final prompt pool size, total trainable prompt parameters, or per-task expansion counts for LDEPrompt versus the fixed-pool baselines (e.g., L2P, DualPrompt). This omission leaves open the possibility that accuracy improvements are driven by increased effective capacity from dual expansion rather than the layer-importance guidance or freezing mechanism, directly undermining attribution of the central claim.

Authors: We agree that these details are necessary for a complete evaluation and to clarify the source of performance gains. In the revised manuscript, we have added Table 4 in Section 4.2 reporting the final prompt pool size, total trainable prompt parameters, and per-task expansion counts for LDEPrompt in comparison to L2P and DualPrompt. We have also expanded the discussion in Section 4.2 to explain that expansion is selective and limited by the freezing of lower-importance prompts, preventing unbounded growth. To further address attribution, we include results from a non-expanding (fixed-pool) variant of LDEPrompt, which exhibits lower accuracy than the full model, indicating that the layer-importance guidance and dual-pool mechanisms contribute beyond raw capacity increases. revision: yes
Referee: [§3.3, §4.3] §3.3 and §4.3: The layer-importance scoring and dual-pool expansion rules are described, yet the ablation studies do not include a controlled comparison that holds total prompt capacity fixed while varying only the guidance and expansion logic. Without such an isolation experiment, it is impossible to confirm that the proposed mechanisms, rather than simply more prompts, produce the reported gains.

Authors: We recognize the importance of isolating the contribution of the guidance and expansion logic. In the revised Section 4.3, we have added a new ablation study that compares the full LDEPrompt to a controlled fixed-capacity variant. In this variant, the prompt pool size is capped at the average size reached by the expandable model, and prompt selection follows a non-guided (random) strategy. The results demonstrate that the layer-importance guided expandable version achieves higher accuracy than the fixed-capacity non-guided counterpart. This provides supporting evidence that the proposed mechanisms offer benefits beyond simply increasing the number of prompts. We have updated the text to clarify this comparison. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper is an empirical method proposal for class-incremental learning via a new prompt pool architecture. It contains no equations, derivations, or first-principles predictions that could reduce to inputs by construction. Claims rest on experimental results on standard benchmarks rather than any self-definitional, fitted-input, or self-citation load-bearing steps. Prior prompt-based CIL work is referenced only as motivation; no uniqueness theorems, ansatzes, or renamings from the same authors are invoked to force the result. The central performance claims are therefore independent of the listed circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review is limited to the abstract; no explicit free parameters, axioms, or invented physical entities are described. The core contribution is a new algorithmic technique rather than new entities or fitted constants.

pith-pipeline@v0.9.0 · 5442 in / 1081 out tokens · 31817 ms · 2026-05-10T15:44:57.693682+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

26 extracted references · 5 canonical work pages · 2 internal anchors

[1]

The ultimate goal is to enable deep learning models to continuously absorb new knowl- edge without significantly increasing resource consumption

INTRODUCTION Incremental learning (also known as continual learning or lifelong learning) aims to acquire new task knowledge with- out forgetting previously learned tasks, while avoiding the need to store old task data [1]. The ultimate goal is to enable deep learning models to continuously absorb new knowl- edge without significantly increasing resource ...
[2]

LDEPrompt: Layer-importance guided Dual Expandable Prompt Pool for Pre-trained Model-based Class-Incremental Learning

PRELIMINARIES 2.1. Class-Incremental Learning Class incremental learning (CIL) aims to build a unified clas- sifier from streaming data. In PTM-based CIL, the focus is primarily on exemplar-free CIL, where only the data from the arXiv:2604.11091v1 [cs.CV] 13 Apr 2026 Fig. 1: Illustration of LDEPrompt. The upper part (from left to right) shows: (1) the for...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[3]

Layer Importance Evaluation Before training, we forward the data of the current taskD t into the pretrained backbone modelf θ

METHOD 3.1. Layer Importance Evaluation Before training, we forward the data of the current taskD t into the pretrained backbone modelf θ. For each layerl, we compute its information gain, defined as the difference be- tween the mutual information of the layer input and output: IG(l) =I(x;h l)−I(x;h l−1),(2) whereh l−1 andh l denote the representations be...
[4]

EXPERIMENTS 4.1. Implementation Details Dataset:We followed the research of [4, 16, 18] and selected five commonly used datasets to evaluate the algorithm’s per- formance: CIFAR100 [9], CUB [10], and VTAB [11]. Specif- ically, CIFAR100 consists of 100 classes, CUB contains 200 classes, and VTAB includes 50 classes. Dataset split:Following the benchmark se...

1993
[5]

CONCLUSION In this work, we presented LDEPrompt, a layer-importance guided dual expandable prompt pool framework designed to alleviate catastrophic forgetting while supporting scal- able prompt-based learning. By leveraging information gain to adaptively select layers for prompt insertion, and by intro- ducing a dual-pool design—consisting of a frozen glo...
[6]

Class-incremental learning: A survey,

Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu, “Class-incremental learning: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

2024
[7]

Catastrophic forgetting in connec- tionist networks,

Robert M French, “Catastrophic forgetting in connec- tionist networks,”Trends in cognitive sciences, vol. 3, no. 4, pp. 128–135, 1999

1999
[8]

Continual learning with pre- trained models: A survey,

Da-Wei Zhou, Hai-Long Sun, Jingyi Ning, Han-Jia Ye, and De-Chuan Zhan, “Continual learning with pre- trained models: A survey,” inIJCAI, 2024, pp. 8363– 8371

2024
[9]

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, 2022, pp. 139–149

2022
[10]

Dualprompt: Complementary prompting for rehearsal-free continual 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 continual learning,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 631–648

2022
[11]

Coda- prompt: Continual decomposed attention-based prompt- ing 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 prompt- ing for rehearsal-free continual learning,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 11909–11919

2023
[12]

Evolving parameterized prompt memory for continual learning,

Muhammad Rifki Kurniawan, Xiang Song, Zhiheng Ma, Yuhang He, Yihong Gong, Yang Qi, and Xing Wei, “Evolving parameterized prompt memory for continual learning,” inProceedings of the AAAI Conference on Artificial Intelligence, 2024, vol. 38, pp. 13301–13309

2024
[13]

Convolutional prompting meets language models for continual learning,

Anurag Roy, Riddhiman Moulick, Vinay K Verma, Sap- tarshi Ghosh, and Abir Das, “Convolutional prompting meets language models for continual learning,” inPro- ceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, 2024, pp. 23616–23626

2024
[14]

Learning multiple layers of features from tiny images,

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

2009
[15]

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

2011
[16]

A large-scale study of representation learning with the visual task adaptation benchmark

Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al., “A large-scale study of repre- sentation learning with the visual task adaptation bench- mark,”arXiv preprint arXiv:1910.04867, 2019

work page arXiv 1910
[17]

A comprehensive survey of continual learning: theory, method and application,

Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu, “A comprehensive survey of continual learning: theory, method and application,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

2024
[18]

Tae: Task-aware expandable representation for long tail class incremental learning,

Linjie Li, Six Liu, Zhenyu Wu, et al., “Tae: Task-aware expandable representation for long tail class incremental learning,”arXiv preprint arXiv:2402.05797, 2024

work page arXiv 2024
[19]

Visual prompt tuning,

Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser- Nam Lim, “Visual prompt tuning,” inEuropean Confer- ence on Computer Vision. Springer, 2022, pp. 709–727

2022
[20]

Prefix-Tuning: Optimizing Continuous Prompts for Generation

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

work page internal anchor Pith review arXiv 2021
[21]

Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need,

Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu, “Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need,”International Journal of Computer Vision, pp. 1–21, 2024

2024
[22]

Consistent prompting for rehearsal-free continual learning,

Zhanxin Gao, Jun Cen, and Xiaobin Chang, “Consistent prompting for rehearsal-free continual learning,” inPro- ceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, 2024, pp. 28463–28473

2024
[23]

Mote: Mixture of task-specific experts for pre-trained model-based class- incremental learning,

Linjie Li, Zhenyu Wu, and Yang Ji, “Mote: Mixture of task-specific experts for pre-trained model-based class- incremental learning,”Knowledge-Based Systems, p. 113795, 2025

2025
[24]

icarl: Incremental classifier and representation learning,

Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert, “icarl: Incremental classifier and representation learning,” inProceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2017, pp. 2001–2010

2017
[25]

Au- tomatic differentiation in pytorch,

Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer, “Au- tomatic differentiation in pytorch,” 2017

2017
[26]

Pilot: A pre-trained model-based contin- ual learning toolbox,

Hai-Long Sun, Da-Wei Zhou, Han-Jia Ye, and De- Chuan Zhan, “Pilot: A pre-trained model-based contin- ual learning toolbox,”arXiv preprint arXiv:2309.07117, 2023

work page arXiv 2023