MePo++: Unifying Representation Refinement and Reconciliation for General Continual Learning

Guanglong Sun; Kanglei Zhou; Liyuan Wang; Qi Cheng; Hongwei Yan; Shuang Cui; Hang Su; Jun Zhu

cs.AISep 4, 2026

MePo++: Unifying Representation Refinement and Reconciliation for General Continual Learning

Authors: Guanglong Sun, Kanglei Zhou, Liyuan Wang, Qi Cheng, Hongwei Yan, Shuang Cui, Hang Su, Jun Zhu,

Organizations: School of Life Sciences, IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing, China · Department of Psychological and Cognitive Sciences, Tsinghua University, Beijing, China · Institute of Software, Chinese Academy of Sciences, Beijing, China · Department of Computer Science and Technology, Institute for AI, BNRist Center, THBI Lab, Tsinghua-Bosch Joint Center for ML, Tsinghua University, Beijing, China

Abstract

General continual learning (GCL) aims to learn from evolving data streams without task identities, explicit boundaries, or repeated access to previous data, making it a realistic yet challenging setting for continual intelligence. Although pretrained models (PTMs) provide rich prior knowledge for addressing the limited supervision and non-stationary nature of GCL, existing PTM-based methods often directly adapt pretrained representations and overlook two critical gaps: the misalignment between upstream pretraining and downstream continual adaptation, and the unreliability of conventional output alignment under blurry streams. Here we propose MePo++, a unified post-training framework that bridges pretrained knowledge and downstream GCL through representation refinement and reconciliation. MePo++ introduces two complementary components: MetaPrep, which improves representation plasticity for continual adaptation through unsupervised meta-refinement over pseudo continual sequences; and StreamAlign, which reinforces representation stability by reconciling evolving online features with a stable pretrained geometry. By improving representation learnability before adaptation and preserving alignment during continual learning, MePo++ enables PTMs to remain both plastic for new concepts and stable over evolving streams. Experiments across diverse PTMs, datasets, and continual learning baselines demonstrate the consistent effectiveness and generality of MePo++ for PTM-based GCL. Our code is available at https://github.com/SunGL001/MePo_Plus.

Figures & tables

Panel of Figure 1 — Fig. 1 : Motivation and performance overview. (a) Existing PTM-based GCL directly adapts pretrained representations to evolving data streams, leading to an upstream-downstream gap and a downstream alignment gap. (b) MePo++ addresses the two gaps through MetaPrep, which refines pretrained representations before deployment, and StreamAlign, which reconciles evolving representations during continual adaptation. (c–e) Performance comparisons under Sup-21K, Sup-21/1K, and DINOv2 pretraining across CIFAR-100, ImageNet-R, and CUB-200, demonstrating consistent improvements over representative GCL baselines across all evaluated settings.

Fig. 2 : Empirical analysis of post-training strategies for improving continual learnability. We report the performance changes over the original DualPrompt with a Sup-21K pretrained backbone, where the baseline results are provided in Tab. I . Results on CIFAR-100, ImageNet-R, and CUB-200 show the effects of different post-training strategies on GCL performance. Positive and negative values indicate improvements and degradations compared with the original baseline, respectively. While existing post-training strategies for downstream adaptation yield limited or unstable gains, MetaPrep consistently improves both

A_{\mathrm{AUC}}

and

A_{\mathrm{Last}}

, suggesting that pretrained representations benefit from dedicated optimization for continual adaptation.

Fig. 3 : Empirical analysis of output alignment strategies. We report the performance improvement over the original DualPrompt baseline with a Sup-21K pretrained backbone, where the baseline results are provided in Tab. I . Results are shown for different output alignment methods on CIFAR-100, ImageNet-R, and CUB-200. While conventional alignment strategies provide limited and dataset-dependent improvements, StreamAlign consistently achieves stronger gains in both

A_{\mathrm{AUC}}

and

A_{\mathrm{Last}}

, demonstrating the effectiveness of maintaining stable representation alignment under GCL.

Fig. 4 : Framework overview of MePo++. MePo++ unifies representation refinement before deployment and representation reconciliation during continual adaptation. In Stage 1, MetaPrep constructs pseudo continual sequences from unlabeled pretraining data and performs bi-level meta-refinement to obtain a GCL-oriented encoder together with a stable representation geometry prior, addressing the upstream-downstream misalignment. In Stage 2, StreamAlign reconciles transient online representations with this stable prior through geometry-guided reconstruction and semantic regularization, maintaining reliable and discriminative representations under evolving GCL streams.

Fig. 5 : Bi-level meta-refinement in MetaPrep. At each meta-epoch, unlabeled upstream samples are clustered into pseudo classes and arranged as a pseudo continual sequence together with a held-out joint evaluation set. Sequential adaptation precedes joint refinement, and their parameter displacement defines the meta-update toward an initialization optimized for continual adaptation.

Fig. 6 : Illustration of representation reconciliation in StreamAlign. The transient pre-aligned representation $\mathbf{f}_{i}$ may preserve strong class separability but exhibit an imbalanced geometry, whereas direct geometric alignment produces a more balanced representation $\hat{\mathbf{f}}_{i}$ at the potential cost of reduced semantic separability. StreamAlign reconciles the two representations into $\mathbf{f}_{\mathrm{trans},i}$ , preserving both geometric balance and discriminative structure for continual prediction.

PTM	Method	CIFAR-100		ImageNet-R		CUB-200
PTM	Method	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$
Sup-21K	Seq FT	$19.71$ $\,\pm\,$ 3.39	$10.42$ $\,\pm\,$ 4.92	$7.51$ $\,\pm\,$ 3.94	$2.29$ $\,\pm\,$ 0.85	$3.47$ $\,\pm\,$ 0.41	$1.49$ $\,\pm\,$ 0.42
	Linear Probe	$49.69$ $\,\pm\,$ 6.09	$23.07$ $\,\pm\,$ 7.33	$29.24$ $\,\pm\,$ 1.26	$16.87$ $\,\pm\,$ 3.14	$28.96$ $\,\pm\,$ 2.46	$17.33$ $\,\pm\,$ 3.08
	Seq FT (SL [ 42 ] )	$64.90$ $\,\pm\,$ 7.18	$62.06$ $\,\pm\,$ 1.89	$47.20$ $\,\pm\,$ 1.47	$39.60$ $\,\pm\,$ 2.43	$56.16$ $\,\pm\,$ 4.32	$56.50$ $\,\pm\,$ 3.08
	MVP [ 5 ]	$68.13$ $\,\pm\,$ 4.34	$60.56$ $\,\pm\,$ 2.57	$41.50$ $\,\pm\,$ 1.15	$34.14$ $\,\pm\,$ 3.95	$56.78$ $\,\pm\,$ 2.88	$50.25$ $\,\pm\,$ 3.53
	CODA-P [ 44 ]	$78.81$ $\,\pm\,$ 3.38	$80.30$ $\,\pm\,$ 1.58	$50.11$ $\,\pm\,$ 2.14	$46.17$ $\,\pm\,$ 2.00	$64.96$ $\,\pm\,$ 3.30	$59.28$ $\,\pm\,$ 3.14

TABLE I : Overall performance comparison of different methods under the GCL setting on 5-task CIFAR-100, ImageNet-R, and CUB-200. All results are averaged over five runs with different task sequences and reported as mean

\pm

standard deviation.

PTM	Method	CIFAR-100		ImageNet-R		CUB-200
PTM	Method	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$
Sup-21K	L2P [ 24 ]	$49.63$ $\,\pm\,$ 2.65	$52.09$ $\,\pm\,$ 2.49	$14.37$ $\,\pm\,$ 0.83	$12.82$ $\,\pm\,$ 0.76	$46.77$ $\,\pm\,$ 1.38	$44.89$ $\,\pm\,$ 1.35
	w/ MePo	$54.54$ $\,\pm\,$ 2.09	$58.94$ $\,\pm\,$ 0.94	$16.56$ $\,\pm\,$ 0.85	$15.06$ $\,\pm\,$ 1.08	$47.63$ $\,\pm\,$ 1.48	$47.74$ $\,\pm\,$ 1.70
	w/ MePo++	58.16 $\,\pm\,$ 2.23	63.74 $\,\pm\,$ 1.46	21.72 $\,\pm\,$ 0.69	19.93 $\,\pm\,$ 0.92	50.56 $\,\pm\,$ 1.37	49.41 $\,\pm\,$ 1.58
	DualPrompt [ 25 ]	$52.13$ $\,\pm\,$ 3.42	$53.57$ $\,\pm\,$ 2.68	$23.35$ $\,\pm\,$ 0.93	$20.22$ $\,\pm\,$ 2.07	$48.86$ $\,\pm\,$ 3.69	$44.36$ $\,\pm\,$ 2.28
	w/ MePo	$54.02$ $\,\pm\,$ 4.76	$59.61$ $\,\pm\,$ 3.51	$22.31$ $\,\pm\,$ 1.26	$19.28$ $\,\pm\,$ 1.62	$48.82$ $\,\pm\,$ 2.75	$46.64$ $\,\pm\,$ 1.37

TABLE II : Overall performance comparison of different methods under the few-shot GCL setting ( 20% train dataset) on 5-task CIFAR-100, ImageNet-R, and CUB-200. All results are averaged over five runs with different task sequences and reported as mean

\pm

standard deviation for each configuration.

Method	CIFAR-100		ImageNet-R
Method	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$
L2P [ 24 ]	$76.34$ $\,\pm\,$ 1.58	$68.23$ $\,\pm\,$ 2.50	$76.51$ $\,\pm\,$ 1.12	$71.14$ $\,\pm\,$ 1.43
w/ MePo++	77.18 $\,\pm\,$ 1.50	69.53 $\,\pm\,$ 2.70	77.24 $\,\pm\,$ 1.28	71.68 $\,\pm\,$ 1.24
DualPrompt [ 25 ]	$71.19$ $\,\pm\,$ 1.01	$65.83$ $\,\pm\,$ 0.19	$73.37$ $\,\pm\,$ 0.77	$66.56$ $\,\pm\,$ 0.11
w/ MePo++	74.27 $\,\pm\,$ 1.07	68.72 $\,\pm\,$ 0.17	76.29 $\,\pm\,$ 0.79	69.90 $\,\pm\,$ 0.18

TABLE III : Performance comparison of prompt-based learners with a CLIP backbone under the Si-Blurry GCL protocol on CIFAR-100 and ImageNet-R. Results are averaged over five random seeds and reported as mean

\pm

standard deviation. Higher values are better for both metrics across all reported datasets.

Setting	DualPrompt				MISA
Setting	+Param.	+Ratio	Time (s)	Accuracy	+Param.	+Ratio	Time (s)	Accuracy
Baseline	637k	0.74%	4.87	55.73	653k	0.76%	4.84	65.40
w/ MePo	1213k	1.41%	4.99	58.36	1229k	1.43%	5.01	68.13
w/ MePo++	1213k	1.41%	5.09	59.05	1229k	1.43%	5.05	69.52

TABLE IV : Computational efficiency comparison on CUB-200 using the Sup-21K pretrained backbone.

MetaPrep	StreamAlign	DualPrompt		MISA
MetaPrep	StreamAlign	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$
✗	✗	$66.36$ $\,\pm\,$ 4.42	$58.09$ $\,\pm\,$ 4.40	$80.35$ $\,\pm\,$ 2.39	$80.75$ $\,\pm\,$ 1.24
✓	✗	$70.65$ $\,\pm\,$ 3.25	$65.27$ $\,\pm\,$ 4.15	$81.29$ $\,\pm\,$ 2.28	$81.84$ $\,\pm\,$ 1.00
✗	✓	$69.10$ $\,\pm\,$ 5.22	$65.76$ $\,\pm\,$ 5.59	$81.36$ $\,\pm\,$ 2.46	$82.19$ $\,\pm\,$ 0.78
✓	✓	72.99 $\,\pm\,$ 4.11	72.00 $\,\pm\,$ 2.85	82.73 $\,\pm\,$ 0.84	84.01 $\,\pm\,$ 0.50

TABLE V : Ablation study of the core modules using Sup-21K.

Post-training Strategy	DualPrompt		MISA
Post-training Strategy	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$
w/o Post-training	$42.47$ $\,\pm\,$ 1.75	$36.92$ $\,\pm\,$ 1.02	$52.61$ $\,\pm\,$ 1.74	$48.01$ $\,\pm\,$ 2.05
w/ Post-training (OML [ 41 ] )	$46.86$ $\,\pm\,$ 2.14	$39.11$ $\,\pm\,$ 1.93	$54.54$ $\,\pm\,$ 2.67	$49.29$ $\,\pm\,$ 0.98
w/ Post-training (MePo [ 15 ] )	$45.48$ $\,\pm\,$ 1.63	$39.09$ $\,\pm\,$ 1.31	$54.10$ $\,\pm\,$ 2.16	$50.05$ $\,\pm\,$ 1.84
w/ Post-training (MePo++)	47.34 $\,\pm\,$ 1.63	40.77 $\,\pm\,$ 1.84	55.18 $\,\pm\,$ 1.72	50.20 $\,\pm\,$ 1.43

TABLE VI : Effect of different representation preparation strategies on ImageNet-R using the Sup-21K pretrained backbone with DualPrompt and MISA.

Output Alignment Strategy	DualPrompt		MISA
Output Alignment Strategy	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$	$A_{\mathrm{AUC}}~(\uparrow)$	$A_{\mathrm{Last}}~(\uparrow)$
w/o Alignment	$43.14$ $\,\pm\,$ 2.18	$34.93$ $\,\pm\,$ 1.50	$43.44$ $\,\pm\,$ 1.80	$33.78$ $\,\pm\,$ 1.63
w/ Mask Alignment (MISA [ 6 ] )	$44.33$ $\,\pm\,$ 2.38	$35.01$ $\,\pm\,$ 1.12	$52.63$ $\,\pm\,$ 1.95	$46.03$ $\,\pm\,$ 1.21
w/ Covariance Alignment (MePo [ 15 ] )	$44.29$ $\,\pm\,$ 2.14	$36.26$ $\,\pm\,$ 1.32	$53.26$ $\,\pm\,$ 2.32	$47.81$ $\,\pm\,$ 1.56
w/ StreamAlign (MePo++)	47.34 $\,\pm\,$ 1.63	40.77 $\,\pm\,$ 1.84	55.18 $\,\pm\,$ 1.72	50.20 $\,\pm\,$ 1.43

TABLE VII : Effect of different representation reconciliation strategies on ImageNet-R using the Sup-21K pretrained backbone with DualPrompt and MISA.

Fig. 7 : Sensitivity analysis of key hyperparameters. (a) Number of MetaPrep training epochs. (b) Number of pseudo-tasks

T^{\prime}

under different stream granularities. (c) Geometric alignment weight

\alpha

. (d) Semantic alignment weight

\lambda

during online adaptation.

Fig. 8 : Representation quality across incremental tasks. t-SNE visualizations of representations learned by MISA and MePo++ from T0 to T4. Marker shapes denote different classes, and higher separation scores indicate better class-level discriminability.

Fig. 9 : Sparse feature activations across incremental tasks. Feature activation patterns of MISA (top) and MePo++ (bottom) from T0 to T4. Activations below $0.7$ are suppressed. MePo++ consistently exhibits lower active rates, indicating sparser representations.

Fig. 10 : Effect of representation reconciliation across incremental tasks. Overlaid t-SNE visualizations of pre-aligned and reconciled features from T0 to T4. Marker shapes denote classes, colors distinguish feature types, and higher separation scores indicate greater class-level discriminability across tasks.

Fig. 11 : Loss landscape comparison between DualPrompt (left) and MePo++ (right). MePo++ exhibits a broader and flatter low-loss region around the converged solution, indicating reduced sensitivity to parameter perturbations during continual adaptation.

Fig. 12 : Case study of prediction outcomes and ground-truth confidence. (a) Overall prediction outcomes of MISA and MePo++; “MePo++ pGT $\uparrow$ ” denotes samples misclassified by both methods for which MePo++ assigns a higher probability to the ground-truth class than MISA. (b–c) Evolution of ground-truth probability for representative classes across evaluation tasks. (d–f) Examples corrected by MePo++ with substantially higher ground-truth confidence. (g) A failure case where both methods are incorrect, while MePo++ still assigns a higher probability to the ground-truth class.

Continual Learning

Continual Learning studies how models acquire knowledge from changing data streams while retaining or selectively revising what they learned before.

See all 301 papers

Aug 17, 2026

Task-Anchored Representation Shaping for Pre-Trained Model-Based Continual Learning

Pre-trained models (PTMs) provide a strong foundation for continual learning by offering stable representations that facilitate lightweight adaptation to new tasks. However, adapting well to each task does not ensure reliable inference over all learned tasks. Since task boundaries are often artificial and semantically entangled, an input from an unknown task can remain ambiguous even with strong PTM features, making cross-task prediction a key bottleneck. We propose Task-Anchored Inference Latent Shaping (TAILS), a lightweight post-PTM module that can be integrated into diverse continual learners and optimized through a decoupled step. TAILS uses fixed task anchors as persistent references to accumulated knowledge. It interprets each sample's feature representation relative to these references, then composes relevant evidence across tasks into latent recall. Rather than selecting a task-specific path or adjusting classifier outputs, TAILS uses latent recall to directly correct the feature representation before prediction. It therefore resolves cross-task ambiguity at the representation level, while leaving the original PTM, method-specific modules, and classifier unchanged. Extensive experiments across multiple PTM-based continual learning paradigms show that TAILS can improve classification and task-inference performance with modest parameter overhead and negligible inference cost.

Zhiming Xu, Huiyu Yi, Zhen-Hao Xie +4

National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China · School of Artificial Intelligence, Nanjing University, Nanjing, China · School of Electronic Science and Engineering, Nanjing University, China +1

Incremental Learning

Ask PDF

Mar 11, 2026

Representation Finetuning for Continual Learning

The world is inherently dynamic, and continual learning aims to enable models to adapt to ever-evolving data streams. While pre-trained models have shown powerful performance in continual learning, they still require finetuning to adapt effectively to downstream tasks. However, prevailing Parameter-Efficient Fine-Tuning (PEFT) methods operate through empirical, black-box optimization at the weight level. These approaches lack explicit control over representation drift, leading to sensitivity to domain shifts and catastrophic forgetting in continual learning scenarios. In this work, we introduce Continual Representation Learning (CoRe), a novel framework that for the first time shifts the finetuning paradigm from weight space to representation space. Unlike conventional methods, CoRe performs task-specific interventions within a low-rank linear subspace of hidden representations, adopting a learning process with explicit objectives, which ensures stability for past tasks while maintaining plasticity for new ones. By constraining updates to a low-rank subspace, CoRe achieves exceptional parameter efficiency. Extensive experiments across multiple continual learning benchmarks demonstrate that CoRe not only preserves parameter efficiency but also significantly outperforms existing state-of-the-art methods. Our work introduces representation finetuning as a new, more effective and interpretable paradigm for continual learning.

Haihua Luo, Xuming Ran, Tommi Kärkkäinen +4

Faculty of Information Technology, University of Jyväskylä, Jyväskylä, Finland · National University of Singapore, Singapore, Singapore · School of Computer Science and Technology, Dalian University of Technology, Dalian, China

Continual Learning Fine-Tuning Parameter-Efficient Fine-Tuning Catastrophic Forgetting

Ask PDF

Aug 13, 2026

Information-Theoretic Decoupled Prompt Tuning for Continual Learning

Continual learning (CL) aims to incrementally acquire knowledge from sequential data while avoiding catastrophic forgetting. Recently, prompt tuning has attracted increasing attention as an efficient approach for adapting pre-trained models to CL tasks. However, existing prompt design paradigms commonly suffer from retrieval dependence and classifier bias, which make model adaptation sensitive to prompt selection and bias predictions toward newly arrived classes. To address these challenges, we propose Decoupled Prompt Tuning for Continual Learning (DPT4CL), which decouples the CLIP textual prompt into a task-shared prompt distribution and class-specific prompts. The task-shared prompt distribution is derived by optimizing an Information Bottleneck objective to facilitate cross-task knowledge transfer and alleviate classifier bias, while class-specific prompts enhance inter-class separability without relying on explicit prompt retrieval. Furthermore, we establish a unified excess risk bound from an information-theoretic perspective, providing theoretical support for the robust generalization and forgetting mitigation of the proposed framework. Extensive experiments on standard CL benchmarks demonstrate that DPT4CL achieves state-of-the-art performance. The source code is available at https://github.com/Cloudfly-Z/DPT4CL

Yunfei Zhang, Wen Wen, Tieliang Gong +1

National Engineering Lab for Big Data Analytics, School of Computer Science and Technology, Xi’an Jiaotong University, Xi’an 710049, China

Continual Learning for VLMs Continual Learning for LLMs VLM Adaptation

Ask PDF

Abstract

Continual Learning

Explore similar work

Task-Anchored Representation Shaping for Pre-Trained Model-Based Continual Learning

Representation Finetuning for Continual Learning

Information-Theoretic Decoupled Prompt Tuning for Continual Learning