Semantic Concept Steering Breaks the Explanation Drift Loop in Continual Learning

Yehonatan Elisha; Oren Barkan; Noam Koenigstein

cs.CVOct 24, 2026

Semantic Concept Steering Breaks the Explanation Drift Loop in Continual Learning

Authors: Yehonatan Elisha, Oren Barkan, Noam Koenigstein

Organizations: Tel Aviv University · The Open University

Abstract

Catastrophic forgetting in continual learning (CL) manifests not only as accuracy degradation but also as explanation drift: models maintain predictive accuracy while silently shifting attention from diagnostic features to spurious correlations. This work identifies and resolves a previously overlooked structural flaw in explanation-aware CL: pixel-level saliency regularization derives supervision targets from the model's own frozen saliency maps, creating a self-referential drift loop that actively reinforces spurious attention across tasks. We introduce \textbf{C $^3$ L} (Concept-Consistent Continual Learning), which breaks this loop by anchoring regularization to class-consistent semantic concepts (e.g., curved beak'', striped pattern'') discovered offline, independent of model state, providing (i) target independence from model drift, and (ii) cross-instance semantic consistency across every instance of a class. We identify that our concept-based objective, combined with approximate LRP relevance conservation, induces a crowding-out effect: since concept and background regions compete for the relevance of the same prediction, enforcing high relevance within concept regions implicitly suppresses spurious background correlations without explicit penalization. \method does not require human-annotated concept sets, as it employs a fully automated concept discovery and spatial grounding. Extensive evaluation across five benchmarks and over 20 baselines demonstrates state-of-the-art performance in accuracy, forgetting, explanation quality, spurious correlation mitigation, and out-of-distribution robustness.

Figures & tables

Figure 1: Mitigating Explanation Drift via Semantic Anchoring. Saliency maps for “Electric Guitar” sample after continual learning. Standard CL (PLAN) suffers from explanation drift, shifting attention to spurious background artifacts. While pixel-level methods (RRR) attempt to mitigate this, they result in noisy, non-specific attributions. Our C 3 L anchors model representations to semantically meaningful concepts, maintaining focus on diagnostic features like the fretboard and guitar body.

Figure 2: C 3 L Overview. Discriminative visual concepts are discovered offline via an LLM and spatially grounded using SAM 3, producing model-state-independent supervision targets that break the self-referential drift loop of prior pixel-level methods. During continual training, $\mathcal{L}_{\text{concept}}$ anchors relevance to semantic regions via a crowding-out effect , while $\mathcal{L}_{\text{stab}}$ preserves background spatial structure against task-specific drift.

ResNet-18
	CUB			DomainNet			ImageNet-R			ImageNet-A			ObjectNet
Method	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
Joint Training	83.04 ${\pm 0.51}$	-	-	74.99 ${\pm 0.22}$	-	-	76.31 ${\pm 0.90}$	-	-	71.22 ${\pm 0.42}$	-	-	73.42 ${\pm 0.91}$	-	-
Fine-Tuning	18.12 ${\pm 0.42}$	31.74 ${\pm 0.81}$	20.63 ${\pm 0.35}$	25.15 ${\pm 0.22}$	20.24 ${\pm 0.18}$	26.51 ${\pm 0.47}$	22.08 ${\pm 0.29}$	26.13 ${\pm 0.67}$	23.35 ${\pm 0.33}$	7.92 ${\pm 0.15}$	30.54 ${\pm 0.92}$	9.94 ${\pm 0.25}$	10.76 ${\pm 0.38}$	37.21 ${\pm 0.74}$	12.25 ${\pm 0.44}$
Replay-based
ER	64.34 ${\pm 0.92}$	14.25 ${\pm 0.58}$	68.42 ${\pm 0.84}$	61.92 ${\pm 0.81}$	8.92 ${\pm 0.55}$	61.99 ${\pm 0.76}$	63.38 ${\pm 0.68}$	11.85 ${\pm 0.49}$	66.15 ${\pm 0.72}$	55.01 ${\pm 0.45}$	19.43 ${\pm 0.66}$	59.80 ${\pm 0.38}$	57.84 ${\pm 0.52}$	17.24 ${\pm 0.58}$	62.01 ${\pm 0.55}$

Table 1: Quantitative comparison of Continual Learning performance across five benchmarks using ResNet-18 and ViT-B architectures. We report Average Accuracy (ACC), Average Forgetting (F), and Average Incremental Accuracy (AIA). We report results over 5 trials.

Table 4

Table 5

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Class	Discovered Concepts
Green Heron	dark glossy cap , long sharp beak , green-grey back , yellow legs , white throat stripe , s-shaped neck, large wingspan
Red-winged Blackbird	red shoulder patches , yellow wing bar , black plumage , black beak , black pointed beak
Golden Retriever	floppy ears , thick tail , golden or cream coat , black nose , friendly expression , wavy fur , athletic build

Appendix

Table 6: Examples of automatically discovered concepts for selected classes, highlighting those that passed the filtering criteria.

Dataset	Generated	Retained	Avg Occurrence	Avg Coverage
CUB	10.3	7.6	0.36	0.51
DomainNet	8.1	5.3	0.39	0.49
ImageNet-R	9.8	6.1	0.32	0.52
ImageNet-A	11.2	6.8	0.29	0.60
ObjectNet	8.5	5.6	0.25	0.64

Appendix

Table 7: Statistical summary of the automated concept discovery and filtering process across all evaluated datasets.

Configuration	ImageNet-R			CUB
Configuration	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
C 3 L	82.74	1.89	83.92	83.65	2.92	84.64
C 3 L -NoReplay	79.82	3.84	78.99	80.94	3.27	81.78
w/o $\mathcal{L}_{\text{stab}}$	78.94	3.01	78.32	78.42	3.86	78.98

Appendix

Table 8: Evaluation on different background design for C 3 L using ViT-B on ImageNet-R and CUB.

Training Stage	$R_{\text{rel}}(C)$ $\uparrow$	$R_{\text{rel}}(B)$ $\downarrow$	CR $\uparrow$	$R_{\text{rel}}(T)$
After $\mathcal{T}_{1}$	0.52	0.46	1.13	0.98
After $\mathcal{T}_{3}$	0.61	0.37	1.65	0.98
After $\mathcal{T}_{5}$	0.65	0.33	1.97	0.98
After $\mathcal{T}_{8}$	0.71	0.28	2.54	0.99
After $\mathcal{T}_{10}$	0.72	0.28	2.57	1.00
Absolute Change	+0.20	-0.18	+1.44	+0.02

Appendix

Table 9: Relevance mass flow analysis across training tasks on CUB

\mathcal{T}_{1}

test set. Results demonstrate reallocation of relevance from background to concept regions, validating the crowding-out mechanism.

	ImageNet-R			CUB
Method	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
ER	72.27 ${\pm 0.82}$	9.13 ${\pm 0.54}$	71.56 ${\pm 0.71}$	70.92 ${\pm 0.94}$	8.48 ${\pm 0.62}$	71.40 ${\pm 0.88}$
C 3 L -NoReplay	79.82 ${\pm 0.65}$	3.84 ${\pm 0.42}$	78.99 ${\pm 0.58}$	80.94 ${\pm 0.72}$	3.27 ${\pm 0.38}$	81.78 ${\pm 0.66}$
C 3 L -FullReplay	80.32 ${\pm 0.58}$	2.96 ${\pm 0.39}$	80.04 ${\pm 0.52}$	80.13 ${\pm 0.64}$	3.42 ${\pm 0.51}$	81.52 ${\pm 0.59}$
C 3 L	$\mathbf{82.74{\pm 0.76}}$	$\mathbf{1.89{\pm 0.33}}$	$\mathbf{83.92{\pm 0.45}}$	$\mathbf{83.65{\pm 0.28}}$	$\mathbf{2.92{\pm 0.24}}$	$\mathbf{84.64{\pm 0.37}}$

Appendix

Table 10: Ablation study disentangling the contribution of experience replay versus semantic regularization on ImageNet-R and CUB (ViT-B).

Explanation Method	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
C 3 L -AttnLRP (Ours)	$82.74\pm 0.76$	$1.89\pm 0.33$	$83.92\pm 0.45$
C 3 L -GradCAM	$80.92\pm 0.84$	$2.43\pm 0.41$	$82.15\pm 0.52$
C 3 L -IIA	$81.85\pm 0.77$	$2.08\pm 0.35$	$83.12\pm 0.46$
C 3 L -T-Attr	$81.47\pm 0.79$	$2.21\pm 0.37$	$82.68\pm 0.48$

Appendix

Table 11: Performance comparison using different explanation methods (AttnLRP, Grad-CAM, IIA, T-Attr) within the C 3 L framework (ViT-B, ImageNet-R).

Method	5 Tasks			10 Tasks			20 Tasks
Method	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
L2P-R	80.52	3.15	82.13	76.82	3.91	77.34	74.12	7.85	75.20
CODA-R	81.45	2.68	82.90	78.54	2.82	79.52	75.30	6.12	76.85
MOS-R	82.37	1.95	83.86	80.92	2.43	81.55	76.45	4.88	78.10
PLAN-R	82.18	2.09	83.49	80.13	2.58	80.89	76.90	4.50	78.55
C 3 L	84.63	1.10	86.15	82.74	1.89	83.92	81.43	2.21	82.38

Appendix

Table 12: Robustness analysis under varying task granularities (5, 10, and 20 split tasks) on ImageNet-R using the ViT-B backbone.

	ImageNet-R			CUB
Mask Type	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
C 3 L	$82.74\pm 0.76$	$1.89\pm 0.33$	$83.92\pm 0.45$	$83.65\pm 0.28$	$2.92\pm 0.24$	$84.64\pm 0.37$
C 3 L (Object segments)	$79.93\pm 0.91$	$2.84\pm 0.41$	$81.06\pm 0.58$	$78.47\pm 1.24$	$4.01\pm 0.66$	$81.64\pm 0.38$

Appendix

Table 13: Performance comparison between using fine-grained concept masks versus coarse object-level segmentation masks (ViT-B).

Grounding Method	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
C 3 L + Grounded SAM	$81.98\pm 0.82$	$2.12\pm 0.38$	$83.25\pm 0.73$
C 3 L	$\mathbf{82.74\pm 0.76}$	$\mathbf{1.89\pm 0.33}$	$\mathbf{83.92\pm 0.45}$

Appendix

Table 14: Impact of the spatial grounding foundation model: comparison between SAM 3 and Grounded SAM on ImageNet-R (ViT-B).

	ImageNet-R			CUB
Configuration	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
C 3 L (no validation)	$81.93\pm 0.84$	$2.17\pm 0.38$	$83.15\pm 0.52$	$82.47\pm 0.89$	$3.24\pm 0.41$	$83.52\pm 0.76$
C 3 L (with validation)	$\mathbf{82.74\pm 0.76}$	$\mathbf{1.89\pm 0.33}$	$\mathbf{83.92\pm 0.45}$	$\mathbf{83.65\pm 0.28}$	$\mathbf{2.92\pm 0.24}$	$\mathbf{84.64\pm 0.37}$
MOS-R (Best baseline)	80.92 ${\pm 0.42}$	2.43 ${\pm 0.28}$	81.55 ${\pm 0.31}$	78.82 ${\pm 0.65}$	3.84 ${\pm 0.29}$	82.52 ${\pm 0.53}$

Appendix

Table 15: Ablation study on the necessity of the concept filtering and validation step across ImageNet-R and CUB.

$\tau_{\text{conf}}$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
0.2	$82.15\pm 0.85$	$2.08\pm 0.61$	$83.24\pm 0.62$
0.3 (our choice)	$\mathbf{82.74\pm 0.76}$	$\mathbf{1.89\pm 0.33}$	$\mathbf{83.92\pm 0.45}$
0.4	$82.43\pm 0.79$	$1.95\pm 0.46$	$83.61\pm 0.48$
0.5	$81.87\pm 0.83$	$2.14\pm 0.79$	$82.98\pm 0.95$

Appendix

Table 16: Sensitivity analysis of the SAM 3 confidence threshold (

\tau_{conf}

) for concept grounding on ImageNet-R (ViT-B).

$\lambda_{\text{concept}}$	$\lambda_{\text{stab}}$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
0.5	0.6	$80.83\pm 0.89$	$2.34\pm 0.41$	$81.34\pm 0.64$
0.8	1.2	$82.74\pm 0.76$	$1.89\pm 0.33$	$83.92\pm 0.45$
1.0	1.5	$82.51\pm 0.81$	$1.95\pm 0.36$	$82.42\pm 0.81$
1.5	2.0	$81.92\pm 0.87$	$2.11\pm 0.38$	$81.85\pm 0.75$
2.0	2.5	$81.24\pm 0.93$	$2.28\pm 0.42$	$81.37\pm 0.53$

Appendix

Table 17: Sensitivity analysis for regularization hyperparameters

\lambda_{concept}

and

\lambda_{stab}

on ImageNet-R (ViT-B).

Method	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
Fine-Tuning	$8.42\pm 0.67$	$35.74\pm 1.12$	$10.15\pm 0.58$
ER	$42.17\pm 1.35$	$18.94\pm 0.92$	$45.83\pm 1.24$
DER++	$48.23\pm 1.18$	$14.62\pm 0.78$	$51.47\pm 1.09$
EPR	$51.02\pm 0.98$	$12.84\pm 0.26$	$53.57\pm 0.75$
RRR	$49.84\pm 1.24$	$13.95\pm 0.82$	$52.93\pm 1.15$
C 3 L	$\mathbf{59.21\pm 1.08}$	$\mathbf{10.74\pm 0.71}$	$\mathbf{60.18\pm 0.97}$

Appendix

Table 18: Continual learning performance on CUB using a non-pretrained ResNet-18 backbone.

ResNet-18
	CUB			DomainNet			ImageNet-R			ImageNet-A			ObjectNet
Method	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
Joint Training	83.04 ${\pm 0.51}$	-	-	74.99 ${\pm 0.22}$	-	-	76.31 ${\pm 0.90}$	-	-	71.22 ${\pm 0.42}$	-	-	73.42 ${\pm 0.91}$	-	-
Fine-Tuning	18.12 ${\pm 0.42}$	31.74 ${\pm 0.81}$	20.63 ${\pm 0.35}$	25.15 ${\pm 0.22}$	20.24 ${\pm 0.18}$	26.51 ${\pm 0.47}$	22.08 ${\pm 0.29}$	26.13 ${\pm 0.67}$	23.35 ${\pm 0.33}$	7.92 ${\pm 0.15}$	30.54 ${\pm 0.92}$	9.94 ${\pm 0.25}$	10.76 ${\pm 0.38}$	37.21 ${\pm 0.74}$	12.25 ${\pm 0.44}$
Regularization-based
EWC	55.15 ${\pm 0.95}$	22.34 ${\pm 0.72}$	60.82 ${\pm 1.05}$	54.40 ${\pm 0.88}$	14.25 ${\pm 0.64}$	54.81 ${\pm 0.92}$	55.24 ${\pm 0.77}$	18.42 ${\pm 0.58}$	58.36 ${\pm 0.83}$	48.04 ${\pm 0.61}$	28.52 ${\pm 1.15}$	52.75 ${\pm 0.55}$	48.42 ${\pm 0.68}$	26.24 ${\pm 0.79}$	52.04 ${\pm 0.73}$

Appendix

Table 19: Quantitative comparison of Continual Learning performance across five benchmarks using ResNet-18 and ViT-B architectures. We report Average Accuracy (ACC), Average Forgetting (F), and Average Incremental Accuracy (AIA). We report results over 5 trials.

Model	Dataset	Metric	Before	After
Model	Dataset	Metric	Before	EPR	TASS	RRR	C 3 L
ResNet-18	ImageNet-A	Acc	8.59	8.49	6.34	7.92	20.15
	ImageNet-A	mIoU	21.32	21.78	20.83	20.39	43.27
	ObjectNet	Acc	21.88	23.08	21.62	21.56	36.93
	ObjectNet	mIoU	35.41	35.27	34.68	34.16	52.91
ViT-B	ImageNet-A	Acc	13.48	15.34	12.91	13.18	30.82

Appendix

Table 20: Out-of-distribution robustness evaluation. We report accuracy (%) and mIoU on ImageNet-A and ObjectNet before and after fine-tuning with C 3 L using a sparse subset. We compare EPR, TASS, RRR, and C 3 L (for ViT-B, the ViT variants EPR-ViT, TASS-ViT, and RRR-ViT).

Method	ACC $\uparrow$	F $\downarrow$	AIA $\uparrow$
ER	44.21 $\pm$ 1.12	21.45 $\pm$ 0.88	46.92 $\pm$ 0.94
DER++	49.82 $\pm$ 0.75	16.32 $\pm$ 0.54	52.14 $\pm$ 0.61
RRR	51.34 $\pm$ 1.04	14.92 $\pm$ 0.91	53.88 $\pm$ 1.02
EPR	52.65 $\pm$ 0.82	13.51 $\pm$ 0.44	55.07 $\pm$ 0.73
C 3 L	59.12 $\pm$ 0.64	10.36 $\pm$ 0.37	62.43 $\pm$ 0.92
C 3 L -PG	58.83 $\pm$ 0.95	10.78 $\pm$ 0.62	62.11 $\pm$ 0.88

Appendix

Table 21: Continual Learning on PartImageNet. Comparison of concept-guided learning against state-of-the-art baselines.

Method	CUB	ImageNet-R
SAM 3 zero-shot (Variant 1, category names)	61.35 $\pm$ 0.68	48.62 $\pm$ 0.43
SAM 3 zero-shot (Variant 2, attribute prompts)	69.23 $\pm$ 0.49	53.16 $\pm$ 0.23
C 3 L (continual learning, ViT-B)	83.65	82.74

Appendix

Table 22: Zero-shot classification accuracy of SAM 3 variants versus C 3 L on CUB and ImageNet-R. Zero-shot evaluation bypasses continual learning entirely; C 3 L results are from the standard 10-task class-incremental protocol (Table 1 ).

Figure 3: Qualitative comparison of explanation maps produced by C 3 L and baseline methods before and after continual learning. Explanation maps produced using ViT-B w.r.t. the classes (top to bottom): ‘lorikeet’, ’usb cable’, ‘common newt, Triturus vulgaris’, and ‘accordion, piano accordion’.

Regularization-Based Continual Learning

See all 19 papers

May 10, 2026

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

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.

Katarzyna Filus, Kamil Faber, Roberto Corizzo +1

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

Ask PDF

Oct 7, 2026

Continual Learning without Continual Training

Continual learning requires models to adapt to new domains and new classes while retaining prior knowledge. Many existing methods rely on continued optimization, using regularization, replay, or parameter expansion to prevent new updates from overwriting previously learned knowledge. Instead, we propose replacing continual training with continual inference: a PFN-based model that is meta-trained, and then frozen, adapting to new classes only by extending an in-context evidence set. Our model, Latent Concept PFN, performs in-context Bayesian inference over a latent concept space that captures semantic structure shared across domains and classes. As each new domain or class arrives, exemplars are added to the memory; adaptation reflects updated posterior beliefs over latent concepts rather than gradient updates. No parameters are changed, reducing forgetting. The same method handles both domain and class incremental continual learning without task identity. Concept annotations are only used during meta-training, acting as a soft anchor on the latent space rather than a fixed bottleneck. Unlike fixed-vocabulary concept methods, the model also handles noisy, ambiguous, or incomplete annotations by combining concept labels with raw input evidence to discover distinctions beyond the predefined concept set. Experiments on class and domain incremental learning datasets demonstrate competitive continual learning performance while learning interpretable latent concepts.

Nikita Narayanan, Ritham Majumdarr, Sonali Parbhoo

Imperial College London

Continual Learning Class-Incremental Learning Domain-Incremental Learning Prior-Data Fitted Networks In-Context Learning

Ask PDF

Apr 16, 2026

CI-CBM: Class-Incremental Concept Bottleneck Model for Interpretable Continual Learning

Catastrophic forgetting remains a fundamental challenge in continual learning, in which models often forget previous knowledge when fine-tuned on a new task. This issue is especially pronounced in class incremental learning (CIL), which is the most challenging setting in continual learning. Existing methods to address catastrophic forgetting often sacrifice either model interpretability or accuracy. To address this challenge, we introduce ClassIncremental Concept Bottleneck Model (CI-CBM), which leverage effective techniques, including concept regularization and pseudo-concept generation to maintain interpretable decision processes throughout incremental learning phases. Through extensive evaluation on seven datasets, CI-CBM achieves comparable performance to black-box models and outperforms previous interpretable approaches in CIL, with an average 36% accuracy gain. CICBM provides interpretable decisions on individual inputs and understandable global decision rules, as shown in our experiments, thereby demonstrating that human understandable concepts can be maintained during incremental learning without compromising model performance. Our approach is effective in both pretrained and non-pretrained scenarios; in the latter, the backbone is trained from scratch during the first learning phase. Code is publicly available at github.com/importAmir/CI-CBM.

Amirhosein Javadi, Tuomas Oikarinen, Tara Javidi +1

Department of Electrical and Computer Engineering, University of California San Diego · Department of Computer Science and Engineering, University of California San Diego · Halıcıoğlu Data Science Institute, University of California San Diego

Class-Incremental Learning Concept Bottleneck Models Continual Learning Interpretable ML Incremental Learning Catastrophic Forgetting

Ask PDF

Abstract

Regularization-Based Continual Learning

Explore similar work

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

Continual Learning without Continual Training

CI-CBM: Class-Incremental Concept Bottleneck Model for Interpretable Continual Learning