Continual Visual Learning under Evolving Semantic Concept Shift

Ismail Lamaakal; Chaymae Yahyati; Yassine Maleh; Khalid El Makkaoui; Ibrahim Ouahbi

cs.CVAug 24, 2026

Continual Visual Learning under Evolving Semantic Concept Shift

Authors: Ismail Lamaakal, Chaymae Yahyati, Yassine Maleh, Khalid El Makkaoui, Ibrahim Ouahbi

Organizations: Department of Computer Science, Faculty of Applied Sciences Nador, Mohammed Premier University, Oujda, Morocco · Laboratory LaSTI, ENSAK, Sultan Moulay Slimane University, Khouribga, Morocco

Abstract

Visual foundation models are commonly adapted under the assumption that the appearance of incoming data may change while the semantic meaning of the prediction task remains fixed. In long-lived visual systems, however, taxonomies, policies, and concept definitions can themselves evolve, causing the same visual evidence to require a different interpretation. We study this setting as evolving semantic concept shift and introduce SemReWrite, a framework for selectively updating obsolete visual--semantic mappings while preserving knowledge that remains valid. SemReWrite represents changes between old and revised semantic specifications, combines semantic discrepancy with sparse revised supervision to localize affected visual regions, and uses an input-dependent low-rank rewriting mechanism together with structured semantic memory, preservation, and obsolete-decision suppression. We further introduce EvoShift-Bench, spanning ImageNet, iNaturalist, CUB-200-2011, and DomainNet, with semantic transitions including class split, merge, boundary revision, insertion, partial redefinition, recurrence, and mixed semantic--appearance shift. To explicitly evaluate selective semantic revision, we introduce Rewrite Accuracy (RA) and Preservation Accuracy (PA) for affected and unaffected regions, respectively, Obsolete Retention (OR) for measuring residual outdated semantic associations, and the Selective Revision Score (SRS), which jointly summarizes rewriting and preservation performance. Experiments show that SemReWrite achieves a stronger balance between learning revised semantics and retaining unaffected knowledge than prompt replacement, conventional fine-tuning, parameter-efficient adaptation, and continual-learning strategies.

Figures & tables

Fig. 1: Overview of the proposed SemReWrite framework. The old semantic specification $\mathcal{S}_{t}$ and the revised specification $\mathcal{S}_{t+1}$ are first encoded to characterize how the meaning of the task has changed. The resulting semantic change is incorporated into a structured semantic concept memory. In parallel, the frozen visual encoder maps each image into the shared representation space. The affected concept localizer combines the specification-level semantic discrepancy with sparse examples labeled under the new specification to estimate the rewrite probability $a(x)$ . Inputs associated with a large rewrite probability activate the selective rewriting pathway, whereas inputs associated with stable semantics are constrained by the preservation pathway. The two branches jointly produce the updated model $M_{t+1}$ , allowing obsolete semantic mappings to be replaced while retaining visual knowledge that remains valid.

Dataset	Classes	Split	Merge	Boundary	Insertion	Recurring	Mixed
ImageNet	1,000	✓	✓	–	✓	✓	✓
iNaturalist	1,000	✓	✓	–	✓	✓	–
CUB-200	200	–	–	✓	–	✓	–
DomainNet	345	✓	✓	–	✓	–	✓

TABLE I: Overview of the principal EvoShift-Bench families.

ID	Shift	Example
S1	Split	Vehicle $\rightarrow$ Car / Truck
S2	Merge	Husky + Malamute $\rightarrow$ Northern dog
S3	Boundary revision	Semantic rule changes
S4	Class insertion	Parent $\rightarrow$ new subclass
S5	Partial redefinition	Subset of concepts changes
S6	Recurring shift	$\mathcal{S}_{1}\!\rightarrow\!\mathcal{S}_{2}\!\rightarrow\!\mathcal{S}_{1}$

TABLE II: Semantic shift protocols in EvoShift-Bench .

Method	Metric	Split	Merge	Boundary	Redefinition	Insertion	Avg.
Old Model	Acc.	65.2 $\pm$ 0.5	69.1 $\pm$ 0.4	61.0 $\pm$ 0.6	63.1 $\pm$ 0.7	58.5 $\pm$ 0.8	63.4
	SRS	65.2 $\pm$ 0.6	68.1 $\pm$ 0.5	62.0 $\pm$ 0.7	63.4 $\pm$ 0.6	59.0 $\pm$ 0.8	63.5
New Text	Acc.	74.6 $\pm$ 0.6	76.8 $\pm$ 0.5	70.5 $\pm$ 0.8	72.4 $\pm$ 0.7	70.8 $\pm$ 0.7	73.0
	SRS	75.4 $\pm$ 0.6	77.2 $\pm$ 0.5	72.0 $\pm$ 0.7	73.8 $\pm$ 0.7	71.9 $\pm$ 0.8	74.1
Full FT	Acc.	82.1 $\pm$ 0.7	83.6 $\pm$ 0.6	80.4 $\pm$ 0.8	81.0 $\pm$ 0.7	78.8 $\pm$ 0.9	81.2
	SRS	81.0 $\pm$ 0.7	82.1 $\pm$ 0.6	79.5 $\pm$ 0.8	80.0 $\pm$ 0.7	77.8 $\pm$ 0.9	80.1

TABLE III: Overall performance under five semantic shift types using four revised examples per changed concept. Acc. measures revised-task accuracy and SRS jointly measures rewriting and preservation.

Method	RA $\uparrow$	PA $\uparrow$	OR $\downarrow$	SRS $\uparrow$
Old Model	51.6 $\pm$ 0.7	94.2 $\pm$ 0.3	41.8 $\pm$ 1.0	66.7 $\pm$ 0.6
New Text	65.8 $\pm$ 0.8	92.5 $\pm$ 0.4	29.7 $\pm$ 0.9	76.9 $\pm$ 0.6
Full FT	84.7 $\pm$ 0.6	80.1 $\pm$ 0.9	13.6 $\pm$ 0.7	82.3 $\pm$ 0.7
LoRA	83.6 $\pm$ 0.5	84.9 $\pm$ 0.7	14.8 $\pm$ 0.6	84.2 $\pm$ 0.5
EWC	73.8 $\pm$ 0.7	92.4 $\pm$ 0.4	27.1 $\pm$ 0.8	82.1 $\pm$ 0.5
LwF	76.2 $\pm$ 0.6	92.1 $\pm$ 0.4	24.6 $\pm$ 0.7	83.4 $\pm$ 0.5

TABLE IV: Revision–preservation analysis. RA evaluates changed regions, PA evaluates stable regions, and OR measures retained obsolete predictions.

Fig. 2: Selective forgetting analysis. (a) RA–PA trade-off; the desired region is the upper right. (b) SRS and obsolete retention (OR). SemReWrite combines stronger rewriting and preservation with lower obsolete retention.

Method	Metric	1	2	4	8	16
New Text	RA	65.8 $\pm$ 0.8	65.8 $\pm$ 0.8	65.8 $\pm$ 0.8	65.8 $\pm$ 0.8	65.8 $\pm$ 0.8
	PA	92.5 $\pm$ 0.4	92.5 $\pm$ 0.4	92.5 $\pm$ 0.4	92.5 $\pm$ 0.4	92.5 $\pm$ 0.4
	SRS	76.9 $\pm$ 0.6	76.9 $\pm$ 0.6	76.9 $\pm$ 0.6	76.9 $\pm$ 0.6	76.9 $\pm$ 0.6
Full FT	RA	74.2 $\pm$ 1.0	80.5 $\pm$ 0.8	84.7 $\pm$ 0.6	88.2 $\pm$ 0.6	90.1 $\pm$ 0.5
	PA	84.8 $\pm$ 0.8	82.8 $\pm$ 0.8	80.1 $\pm$ 0.9	77.0 $\pm$ 1.0	74.2 $\pm$ 1.1
	SRS	79.1 $\pm$ 0.8	81.6 $\pm$ 0.7	82.3 $\pm$ 0.7	82.2 $\pm$ 0.8	81.4 $\pm$ 0.9

TABLE V: Effect of revised supervision. Each column gives the number of newly labeled examples per changed concept.

Fig. 3: Annotation efficiency. (a) SRS versus revised supervision. (b) RA and PA for SemReWrite and Full FT. SemReWrite improves rewriting while keeping preservation nearly stable.

Method	Metric	Sketch	Corruption	Clipart	Low-light	Avg.
New Text	Acc.	58.3 $\pm$ 0.9	62.1 $\pm$ 0.8	60.7 $\pm$ 0.9	63.8 $\pm$ 0.8	61.2
	SRS	50.2 $\pm$ 1.0	55.1 $\pm$ 0.9	52.4 $\pm$ 0.9	56.3 $\pm$ 0.8	53.5
Full FT	Acc.	69.5 $\pm$ 0.8	72.4 $\pm$ 0.7	71.1 $\pm$ 0.8	73.0 $\pm$ 0.7	71.5
	SRS	63.8 $\pm$ 0.9	67.2 $\pm$ 0.8	65.7 $\pm$ 0.8	68.0 $\pm$ 0.7	66.2
LoRA	Acc.	71.2 $\pm$ 0.7	73.7 $\pm$ 0.6	72.8 $\pm$ 0.7	74.4 $\pm$ 0.6	73.0
	SRS	66.0 $\pm$ 0.8	68.9 $\pm$ 0.7	67.8 $\pm$ 0.7	69.5 $\pm$ 0.6	68.0

TABLE VI: Combined semantic and covariate shift. Acc. evaluates the revised task and SRS additionally evaluates preservation.

Fig. 4: Mixed-shift comparison. (a) Revised-task accuracy and (b) SRS across four visual shifts. SemReWrite remains effective when semantic and appearance changes occur jointly.

Step	Acc.	RA	PA	SRS	OR
$0$	90.6 $\pm$ 0.3	–	94.1 $\pm$ 0.3	–	–
$1$	88.7 $\pm$ 0.4	89.2 $\pm$ 0.4	93.5 $\pm$ 0.3	91.3 $\pm$ 0.3	9.8 $\pm$ 0.5
$2$	88.0 $\pm$ 0.4	88.5 $\pm$ 0.4	93.2 $\pm$ 0.3	90.8 $\pm$ 0.3	9.2 $\pm$ 0.5
$3$	87.4 $\pm$ 0.5	88.0 $\pm$ 0.4	92.9 $\pm$ 0.4	90.4 $\pm$ 0.4	8.8 $\pm$ 0.4
$4$	86.9 $\pm$ 0.5	87.4 $\pm$ 0.5	92.6 $\pm$ 0.4	89.9 $\pm$ 0.4	8.4 $\pm$ 0.4
$5$	86.5 $\pm$ 0.5	86.9 $\pm$ 0.5	92.4 $\pm$ 0.4	89.6 $\pm$ 0.4	8.1 $\pm$ 0.4

TABLE VII: Sequential adaptation across five semantic revisions without model reset.

Fig. 5: Long-term semantic adaptation. (a) Current accuracy, PA, and OR over five revisions. (b) Accuracy before and after returning to

\mathcal{S}_{1}

; smaller recovery gaps indicate better semantic recurrence.

Variant	RA $\uparrow$	PA $\uparrow$	OR $\downarrow$	SRS $\uparrow$
No localization	87.7 $\pm$ 0.6	82.2 $\pm$ 0.8	10.8 $\pm$ 0.6	84.9 $\pm$ 0.6
No $\mathcal{L}_{\mathrm{pres}}$	90.2 $\pm$ 0.5	79.0 $\pm$ 1.0	7.3 $\pm$ 0.5	84.2 $\pm$ 0.7
No $\mathcal{L}_{\mathrm{erase}}$	84.5 $\pm$ 0.6	93.2 $\pm$ 0.4	18.9 $\pm$ 0.8	88.6 $\pm$ 0.5
No semantic graph	86.1 $\pm$ 0.5	92.0 $\pm$ 0.4	10.7 $\pm$ 0.6	89.0 $\pm$ 0.4
No semantic delta	83.2 $\pm$ 0.7	91.8 $\pm$ 0.5	14.9 $\pm$ 0.7	87.3 $\pm$ 0.5
Full-backbone update	89.4 $\pm$ 0.5	84.1 $\pm$ 0.8	8.2 $\pm$ 0.5	86.6 $\pm$ 0.6

TABLE VIII: Ablation of SemReWrite components. Higher RA, PA, and SRS are better, while lower OR is better.

Fig. 6: Ablation and sensitivity analysis. (a) SRS after removing individual components. (b) Effect of preservation weight

\lambda_{p}

. (c) Diagnostic hard-routing threshold

\tau_{a}

; the default method uses the continuous gate.

Failure condition	Reason
Ambiguous update	The revised semantic specification does not clearly define what should change.
Non-visual semantics	The new distinction depends on information that cannot be inferred from the image.
Global revision	Most of the visual space becomes affected, reducing the advantage of selective adaptation.
Subtle change with few labels	Sparse supervision is insufficient to localize a fine-grained revised boundary.
Contradictory update	The new semantic definition conflicts with relations already represented in the concept graph.

TABLE IX: Principal failure modes of SemReWrite and their underlying causes.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Acc.	SRS	Acc.	SRS	Acc.	SRS	Acc.	SRS
	ImageNet		iNaturalist		CUB-200		DomainNet
Full FT	82.1 $\pm$ 0.7	80.4 $\pm$ 0.7	81.1 $\pm$ 0.8	79.5 $\pm$ 0.8	80.4 $\pm$ 0.8	79.0 $\pm$ 0.8	71.5 $\pm$ 0.8	66.2 $\pm$ 0.9
LoRA	83.4 $\pm$ 0.5	83.0 $\pm$ 0.5	82.2 $\pm$ 0.6	82.0 $\pm$ 0.6	81.6 $\pm$ 0.7	81.0 $\pm$ 0.7	73.0 $\pm$ 0.6	68.0 $\pm$ 0.7
LwF	79.8 $\pm$ 0.7	80.5 $\pm$ 0.6	78.6 $\pm$ 0.8	79.2 $\pm$ 0.7	76.6 $\pm$ 0.9	78.0 $\pm$ 0.8	68.9 $\pm$ 0.9	63.4 $\pm$ 0.9
StatA	78.3 $\pm$ 0.8	79.0 $\pm$ 0.7	77.6 $\pm$ 0.8	78.2 $\pm$ 0.8	74.1 $\pm$ 1.0	76.5 $\pm$ 0.9	73.4 $\pm$ 0.6	68.5 $\pm$ 0.7
SemReWrite	88.7 $\pm$ 0.4	91.4 $\pm$ 0.3	87.9 $\pm$ 0.5	90.5 $\pm$ 0.4	87.1 $\pm$ 0.5	90.1 $\pm$ 0.4	79.4 $\pm$ 0.5	76.3 $\pm$ 0.6

Appendix

TABLE X: Per-dataset comparison under the default four-shot setting. Acc. denotes revised-task accuracy and SRS measures the joint revision–preservation trade-off.

Fig. 7: Additional semantic-revision results. (a) Revised accuracy across benchmark families. (b) Accuracy loss after returning to a previously active semantic specification. (c) Obsolete-rule retention across semantic shift types.

Backbone	Method	RA	PA	SRS
CLIP	LoRA	83.6 $\pm$ 0.5	84.9 $\pm$ 0.7	84.2 $\pm$ 0.5
	LwF	76.2 $\pm$ 0.6	92.1 $\pm$ 0.4	83.4 $\pm$ 0.5
	SemReWrite	88.9 $\pm$ 0.4	93.4 $\pm$ 0.3	91.1 $\pm$ 0.3
SigLIP	LoRA	82.8 $\pm$ 0.6	84.0 $\pm$ 0.6	83.4 $\pm$ 0.5
	LwF	75.5 $\pm$ 0.7	91.3 $\pm$ 0.5	82.7 $\pm$ 0.5
	SemReWrite	88.0 $\pm$ 0.4	92.0 $\pm$ 0.4	89.9 $\pm$ 0.3

Appendix

TABLE XI: Generalization across foundation-model backbones. Results are averaged over the five principal semantic shifts. Higher is better.

Fig. 8: SRS across semantic shifts with (a) CLIP, (b) SigLIP, and (c) DINOv2. The advantage of selective rewriting remains consistent across representation families.

Shots	AUROC	AUPRC	F1
1	82.0 $\pm$ 0.7	76.8 $\pm$ 0.9	71.9 $\pm$ 1.0
2	86.3 $\pm$ 0.6	82.1 $\pm$ 0.8	78.4 $\pm$ 0.8
4	90.2 $\pm$ 0.5	87.9 $\pm$ 0.6	84.6 $\pm$ 0.6
8	92.6 $\pm$ 0.4	90.8 $\pm$ 0.5	87.6 $\pm$ 0.5
16	94.0 $\pm$ 0.3	92.5 $\pm$ 0.4	89.4 $\pm$ 0.4

Appendix

TABLE XII: Affected-region localization versus revised supervision. AUROC, AUPRC, and F1 are computed against the benchmark affected-region annotations, which are never used during adaptation.

Fig. 9: Localization and semantic-shift severity. (a) Localization quality versus revised supervision. (b) Relation between localization F1 and SRS. (c) SRS as the affected fraction increases.

Severity	LwF	LoRA	SemReWrite
0.10	84.0	86.2	92.0
0.20	83.7	85.4	91.6
0.30	83.0	84.2	91.1
0.40	81.9	82.9	90.5
0.50	80.4	81.3	89.8
0.60	78.6	79.4	88.9

Appendix

TABLE XIII: SRS under increasing semantic-shift severity. Severity denotes the fraction of evaluation samples whose semantic label changes.

Specification	RA	PA	SRS
Canonical	88.9	93.4	91.1
Paraphrase-1	88.5	93.2	90.8
Paraphrase-2	88.1	93.0	90.5
Short definition	87.4	92.6	89.9
Verbose definition	87.9	92.8	90.3
Ambiguous definition	77.6	88.0	82.4

Appendix

TABLE XIV: Robustness to semantic-description wording. Equivalent paraphrases preserve the intended concept meaning; the ambiguous condition intentionally removes important defining information.

Fig. 10: Robustness to imperfect semantic supervision. (a) Prompt paraphrasing, (b) revised-label noise, and (c) corrupted semantic graph relations.

Label noise	LoRA	SemReWrite
0%	84.2	91.1
5%	83.4	90.7
10%	82.2	89.9
20%	79.8	88.3
30%	76.9	85.9

Appendix

TABLE XV: SRS under noise in the revised few-shot labels. Noise is introduced only into the adaptation set.

Corruption	RA	PA	SRS
0%	88.9	93.4	91.1
10%	88.6	92.9	90.7
20%	88.1	92.0	90.0
30%	87.2	90.7	88.9
40%	86.0	88.9	87.4

Appendix

TABLE XVI: Effect of semantic-graph corruption. The corruption rate is the fraction of graph edges replaced by incorrect same-type relations.

$r$	$2dr$	RA	PA	SRS
2	2,048	84.8	93.7	89.0
4	4,096	87.4	93.6	90.4
8	8,192	88.9	93.4	91.1
16	16,384	89.5	93.0	91.2
32	32,768	89.7	92.6	91.1

Appendix

TABLE XVII: Effect of low-rank rewriting capacity. Adapter parameters count only

U

and

V

for

d=512

; additional localizer and semantic-memory parameters are unchanged across rows.

Fig. 11: Capacity and long-horizon behavior. (a) SRS and PA versus adapter rank. (b) SRS versus low-rank parameter count. (c) SRS across twenty successive semantic revisions.

Transitions	LwF	LoRA	SemReWrite
1	83.5	84.4	91.2
5	82.3	83.0	90.4
10	80.7	81.2	89.4
20	77.6	77.6	87.4

Appendix

TABLE XVIII: SRS after increasingly long semantic sequences. Each value is measured after completing the indicated number of revisions.

Method	First $\mathcal{S}_{1}$	Return $\mathcal{S}_{1}$	Gap $\downarrow$
Full FT	87.9	75.5	12.4
LwF	84.8	81.9	2.9
Replay	85.9	83.1	2.8
SemReWrite	89.0	87.8	1.2

Appendix

TABLE XIX: Recovery after semantic recurrence. The recovery gap is the accuracy decrease between the first and second occurrence of the same semantic specification; lower is better.

Method	Split	Merge	Boundary	Redef.	Insertion
Full FT	12.4	11.8	14.9	13.7	15.1
LoRA	13.8	13.1	16.0	15.2	16.1
LwF	23.4	21.8	27.5	25.7	24.6
Replay	20.7	19.6	23.9	22.8	22.5
SemReWrite	7.3	6.8	8.9	7.8	8.7

Appendix

TABLE XX: Obsolete-rule retention across semantic shift types. Lower values indicate more complete removal of invalid previous decisions.

Appearance shift	RA	PA	OR	SRS
Real $\rightarrow$ Sketch	71.9	76.4	12.8	74.1
Clean $\rightarrow$ Corruption	75.0	79.5	10.9	77.2
Real $\rightarrow$ Clipart	73.8	78.1	11.7	75.9
Normal $\rightarrow$ Low-light	76.0	80.1	10.2	78.0

Appendix

TABLE XXI: Detailed SemReWrite performance under mixed semantic and appearance shift. Higher RA and PA are better; lower OR is better.

Loc.	$\mathcal{L}_{pres}$	$\mathcal{L}_{erase}$	RA	PA	OR	SRS
–	–	–	86.8	76.5	16.2	81.3
✓	–	–	88.0	82.0	14.3	84.9
✓	✓	–	84.5	93.2	18.9	88.6
✓	–	✓	90.2	79.0	7.3	84.2
–	✓	✓	86.9	84.1	9.9	85.5
✓	✓	✓	88.9	93.4	7.9	91.1

Appendix

TABLE XXII: Interaction between localization, preservation, and obsolete decision suppression. A checkmark denotes that the component is active.

Parameter	Default value
Primary backbone	CLIP ViT-B/16
Image resolution	$224\times 224$
Revised labels / concept	4
Adapter rank $r$	8
Localizer hidden dimension	256
Localizer dropout	0.1

Appendix

TABLE XXIII: Default SemReWrite configuration used in the main experiments.

Continual Learning

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

See all 307 papers

Lost or Hidden? A Concept-Level Forgetting in Supervised Continual Learning
May 10, 2026Katarzyna Filus, Kamil Faber, Roberto Corizzo +1Visual Representation Learning Continual Learning
Institute of Theoretical and Applied Informatics Polish Academy of Sciences, Gliwice, Poland · AGH University of Krakow Krakow, Poland · American University Washington DC, USA +1
Provenance Guided Incremental Learning Under Evolving Concept Definitions
Aug 24, 2026Ismail LamaakalIncremental Learning Concept Drift
Department of Computer Science, Faculty of Applied Sciences Nador, Mohammed Premier University, Oujda, Morocco
Towards Continual Test-Time Adaptation of Vision-Language Models in Open-Vocabulary Semantic Segmentation
Aug 30, 2026Chandler Timm C. Doloriel, Yunbei Zhang, Sarthak Kumar Maharana +5Test-Time Adaptation of VLMs Open-Vocabulary Semantic Segmentation
Faculty of Science and Technology (REALTEK), Norwegian University of Life Sciences (NMBU) · Tulane University · University of Texas at Dallas

May 10, 2026J/K move · Enter open · S save

Lost or Hidden? A Concept-Level Forgetting in Supervised Continual Learning

Katarzyna Filus, Kamil Faber, Roberto Corizzo, Christopher Kanan

Institute of Theoretical and Applied Informatics Polish Academy of Sciences, Gliwice, Poland · AGH University of Krakow Krakow, Poland · American University Washington DC, USA+1

Visual Representation Learning Continual Learning Catastrophic Forgetting

Open paper AskPDF arXiv

Continual learning studies how models can adapt to new tasks while retaining previously acquired knowledge. Although a broad spectrum of methods has been proposed to mitigate catastrophic forgetting, the field remains predominantly performance-driven, with limited insight into what forgetting actually corresponds to within the vision model's representation space. Prior work has primarily analyzed forgetting through task-level performance or coarse measures of representational drift, without disentangling output-level accessibility from changes in finer-grained internal structure. To this end, we propose a diagnostic framework that leverages Sparse Autoencoders (SAEs) to define a task-anchored latent feature space, enabling analysis of how task-specific information evolves at a finer granularity, where individual SAE latents are treated as concept proxies for recurring and relatively disentangled visual patterns in the model's internal computations. Within this framework, we decompose forgetting into apparent concept deletion, recoverability, and decodability. We show that a large portion of seemingly lost concept-level information can often be recovered under linearity assumption, with concept decodability degrading as more tasks are introduced. Overall, our findings suggest that a significant part of concept-level forgetting can be attributed to changes in the representational accessibility rather than complete information erasure.

Abstract

Continual Learning

Explore similar work

Lost or Hidden? A Concept-Level Forgetting in Supervised Continual Learning

Provenance Guided Incremental Learning Under Evolving Concept Definitions

Towards Continual Test-Time Adaptation of Vision-Language Models in Open-Vocabulary Semantic Segmentation