Pith. sign in

REVIEW 3 major objections 7 minor 67 references

CL-Anomaly keeps private LoRA experts isolated while layer-adaptive shared experts transfer anomaly knowledge, so one MLLM can keep learning new AD tasks without semantic entanglement.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 06:01 UTC pith:DDB474QM

load-bearing objection Solid first cut at continual MLLM anomaly detection: isolation–sharing plus cheap layer-adaptive expert merge works, with one acknowledged routing soft spot. the 3 major comments →

arxiv 2607.02930 v1 pith:DDB474QM submitted 2026-07-03 cs.CV

CL-Anomaly: Layer-Adaptive Mixture-of-Experts with Multimodal Large Language Model for Continual Learning in Anomaly Detection

Wen Dong , Zhao Wang , Shuangqing Zhang , Kai Sun , Ben Li , Guo-Sen Xie , Caifeng Shan , Fang Zhao This is my paper
classification cs.CV
keywords anomaly detectioncontinual learningmultimodal large language modelmixture-of-expertsLoRAparameter-efficient fine-tuningcatastrophic forgettinglayer-adaptive knowledge transfer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that continual anomaly detection with multimodal large language models is blocked by a double bind: fully shared parameters entangle anomaly knowledge across heterogeneous modalities, domains, and defect scales, while fully private parameters block useful transfer. CL-Anomaly resolves this with an isolation-sharing dual path. Each task gets a frozen-after-training PrivLoRA expert whose parameters never mix with other tasks, so semantic interference cannot occur. In parallel, a small bank of shared LoRA experts lives in a unified space and is updated by Layer-Adaptive Knowledge Transfer: gating probabilities automatically mark the layer-wise key experts for the current task, then a momentum merge protects those key weights while non-key knowledge fades. The result is a single parameter-efficient MLLM that sequentially absorbs class-incremental, cross-domain (industrial to medical), and cross-modal (2D to 3D) anomaly tasks and still approaches the performance of joint multi-task training. A sympathetic reader cares because real AD systems accumulate new products, new imaging modalities, and privacy-constrained data; retraining from scratch is impractical, and prior continual-learning recipes either forget or refuse to transfer.

Core claim

An isolation-sharing dual-path MoE of LoRA experts, combined with zero-cost layer-wise key-expert selection and momentum merging, lets an MLLM continually learn anomaly detection tasks while preventing semantic entanglement and still enabling positive forward and backward knowledge transfer, outperforming prior PEFT continual-learning methods across class-incremental, cross-domain and cross-modal regimes.

What carries the argument

Layer-Adaptive Knowledge Transfer (LAKT): accumulate per-layer gating probabilities to identify the top-K key shared experts for the current task, partition experts into key-only-current, key-only-previous and rest, then apply a three-valued momentum coefficient so key knowledge is preferentially retained while non-key knowledge fades.

Load-bearing premise

At inference the correct private expert is chosen by cosine similarity between the input image’s visual anchor and stored task anchors; if that routing is wrong the isolation path fails.

What would settle it

Replace the visual-anchor router with an oracle that always activates only the true task’s PrivLoRA (or measure routing accuracy on a held-out set of visually similar but semantically different tasks) and check whether the reported Avg/Last/BWT gains collapse or remain.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A single frozen-backbone MLLM can be kept current on new product lines, medical imaging domains and 3D modalities without joint retraining or data replay.
  • Class-incremental industrial inspection can improve previously learned categories as later categories are added (positive backward transfer).
  • Cross-domain jumps (industrial o medical) and cross-modal jumps (2D o multi-view 3D) no longer force a choice between forgetting and isolation.
  • Anomaly reasoning quality (ROUGE-L, SBERT, LLM-Score) after multi-stage training stays competitive with joint multi-task upper bounds.

Where Pith is reading between the lines

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

  • If visual-anchor routing remains the bottleneck, replacing it with a lightweight text-or-instruction-conditioned router could raise the private-path ceiling without increasing storage.
  • The same layer-adaptive key-expert idea may transfer to other heterogeneous continual instruction-tuning settings (e.g., medical report generation or multi-sensor robotics) where domain gaps are large.
  • Because only ~1.4 % extra parameters are added, the method is a practical candidate for edge or on-prem AD deployments that cannot host multiple full models.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes CL-Anomaly, a parameter-efficient continual-learning framework for anomaly detection with MLLMs. It combines task-private PrivLoRA experts (physically isolated LoRA adapters selected by visual-anchor cosine routing) with a shared MoE-LoRA path whose layer-wise key experts are identified from gating statistics and updated by momentum merging (Layer-Adaptive Knowledge Transfer, LAKT). The dual-path design is intended to avoid semantic entanglement while still allowing positive transfer. Experiments on Anomaly-Instruct-125K cover class-incremental (WebAD subsets), cross-domain (industrial→medical BMAD), and cross-modal (2D→3D) sequences, plus a traditional single-category CIL setting on MVTec-AD. Against LoRA, MoELoRA, CL-MoE and HiDe-LLaVA, CL-Anomaly reports large gains in Avg/Last/BWT (Table 1) and in Accuracy/Precision/Recall/F1 (Table 2), with ablations (Table 3), hyper-parameter sweeps (λ, β, expert count), task-order randomization, and qualitative reasoning metrics.

Significance. If the reported gains hold under multi-seed evaluation, the work is a useful and timely contribution: it is among the first systematic treatments of continual MLLM-based anomaly detection under the triple heterogeneity of modality, domain and defect scale, and it supplies a concrete isolation–sharing recipe that approaches multi-task joint training while remaining parameter-efficient (~1.4% overhead). Strengths include a clear dual-path architecture, zero-extra-cost key-expert selection (Eq. 8), multi-scenario evaluation with both CL metrics and AD-appropriate F1-style metrics, component ablations, and an explicit limitation analysis of PrivLoRA routing in Appendix E. Code is promised at a public repository. These elements make the paper of practical interest to industrial/medical AD and to the broader continual-instruction-tuning community.

major comments (3)
  1. [§4.2 Eqs. (3)–(5); Appendix E] Section 4.2, Eqs. (3)–(5) and Appendix E: the isolation path depends on cosine similarity between an instance visual anchor and stored task anchors to weight PrivLoRAs. Appendix E shows that tasks with similar visual distributions (same WebAD source; 3D multi-view tasks) form high-response clusters, so routing is imperfect; an oracle one-hot routing still improves Last accuracy. Because PrivLoRA is load-bearing in the dual-path claim (Table 3: w/o PrivLoRA drops Avg 86.34→83.58 and BWT −0.96→−3.70), the main text should (i) report routing accuracy / confusion across the three scenarios, (ii) quantify how much of the Table 1 margin over CL-MoE survives under noisy routing, and (iii) either propose a lightweight fix (e.g., text+vision anchors, temperature calibration) or clearly bound the regimes where private experts help. Leaving this only in the appendix understates a central assumption
  2. [Table 1; Table 3; Table 7] Tables 1–2 and Fig. 4 report single-run point estimates with no standard deviations, no multi-seed averages, and no statistical tests, yet the central claim is a multi-point margin over CL-MoE (+5.68 Avg / +7.44 Last / +4.94 BWT). Continual-learning rankings are known to be seed- and order-sensitive (the paper itself shows order effects in Table 7). At least 3 seeds (or bootstrap over task orders) for the main Avg/Last/BWT comparison, and error bars on the ablation in Table 3, are needed before the SOTA claim can be treated as settled for a journal audience.
  3. [§5.1; Appendix C.2–C.3] Section 5.1 / Appendix C.2: baseline expert budgets are matched at a total of 12 experts for MoELoRA/CL-MoE, while CL-Anomaly uses 4 shared + one private per task (growing with T). For T=8 this is comparable, but the private experts are never updated after their task and are selected by a separate router; a controlled comparison that freezes the same total trainable parameter count (or the same number of active experts at inference) would strengthen the claim that the isolation–sharing design, rather than extra capacity, drives the gains. Please add a capacity-matched row or a brief parameter-matched ablation.
minor comments (7)
  1. [§4.1 Eq. (2); Fig. 6] Eq. (2) fixes λ=0.5; Fig. 6 shows sensitivity. State the default explicitly in the method section and note that λ is not learned.
  2. [Fig. 2; §4.3] Fig. 2 heatmaps motivate layer-wise preferences but are only qualitative. A short quantitative summary (e.g., entropy of expert usage per layer, or overlap of top-1 experts across tasks) would better justify LAKT.
  3. [§5.1 Evaluation Metrics] Appendix C.1 correctly explains why AUROC is not used (discrete Yes/No outputs). A one-sentence pointer in the main Evaluation Metrics paragraph would help AD readers who expect AUROC.
  4. [Abstract; throughout] Several extracted passages show missing spaces (e.g., abstract “asreal-worldknowledgeevolves”). Clean typesetting and consistent hyphenation of “Layer-Adaptive” / “task-private” throughout.
  5. [Table 4; Appendix D.2] Table 4 LLM-Score uses Qwen-3-Max as judge; report the exact prompt (already in Appendix) and, if possible, inter-judge agreement or a second judge to reduce single-model bias.
  6. [§4.2; Appendix B] Clarify whether visual anchors (Eq. 3) are computed once after each task and frozen, or refreshed; and whether 3D multi-view images use mean pooling over views for the anchor.
  7. [§2] Related Work could briefly position against recent continual AD prompt methods (e.g., UCAD, IUF) beyond the MLLM baselines, even if they are not MLLM-based, to help AD specialists.

Circularity Check

0 steps flagged

No significant circularity: empirical PEFT method whose claims rest on held-out sequential-training metrics, not on any equation that reduces a prediction to its own fitted inputs.

full rationale

CL-Anomaly is a parameter-efficient fine-tuning architecture (PrivLoRA isolation path + Layer-Adaptive Shared Experts + momentum LAKT) whose sole load-bearing claims are comparative numbers (Avg/Last/BWT, Accuracy/Precision/Recall/F1) measured after sequential training on held-out splits of Anomaly-Instruct-125K subsets. Equations 2–10 define the dual-path forward pass, visual-anchor routing, gating, and layer-wise momentum merge; none of them is later re-used as a “prediction” of a quantity already fixed by construction. Hyper-parameters (λ=0.5, β=0.75, n=4, k=1, r=36) are chosen by explicit sweeps (Figs. 6–7, Table 6) and reported, not hidden inside the claim. Ablations (Table 3) and the authors’ own Appendix E routing analysis further expose, rather than conceal, the method’s assumptions. Self-citations to the authors’ prior AD work appear only as ordinary related-work pointers and do not supply a uniqueness theorem, ansatz, or fitted constant that forces the reported ranking. Consequently the derivation chain contains no self-definitional step, no fitted-input-called-prediction, and no load-bearing self-citation circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on a small set of architectural choices and hyper-parameters that are free or domain-assumed rather than derived. No new physical entities are postulated; the invented components are algorithmic modules whose value is measured by the reported tables.

free parameters (4)
  • λ (private/shared path weight) = 0.5
    Fixed at 0.5 after a sweep (Fig. 6); directly scales the two paths in Eq. 2 and therefore affects every reported accuracy.
  • β (momentum merge coefficient for key experts) = 0.75
    Set to 0.75 after a sweep (Fig. 7); controls how much new-task vs old-task knowledge is retained in Eq. 10.
  • n (number of shared experts) and k (key experts per layer) = n=4, k=1
    Chosen as ⟨4,1⟩; Table 6 shows sensitivity; the selection of which experts are “key” depends on this choice.
  • LoRA rank r and scaling α = r=36, α=72
    r=36, α=72; matched across all baselines but still free design choices that set capacity.
axioms (4)
  • domain assumption Layer-wise LoRA experts exhibit distinct task preferences that can be recovered from accumulated gating probabilities without extra forward/backward passes.
    Stated as an observation from Fig. 2 and used to justify the entire LAKT procedure (Eqs. 8–10).
  • domain assumption Visual anchors extracted by the frozen LLaVA encoder are sufficiently discriminative to route an input to the correct PrivLoRA expert via cosine similarity.
    Eqs. 3–5; the paper’s own Appendix E shows this is only approximately true for similar-looking tasks.
  • domain assumption Historical data cannot be replayed (privacy constraint typical of real AD systems), so parameter isolation plus momentum merge must suffice.
    Motivation section and related-work discussion of replay-based methods.
  • standard math Standard LoRA low-rank update form (Eq. 1) and frozen backbone preserve pre-trained knowledge.
    Taken from Hu et al. (LoRA) and used throughout.
invented entities (2)
  • PrivLoRA (task-private LoRA expert) no independent evidence
    purpose: Physically isolate each task’s parameter subspace to prevent semantic entanglement.
    New named module; independent evidence is only the ablation drop when it is removed (Table 3).
  • Layer-Adaptive Knowledge Transfer (LAKT) / Layer-Adaptive Shared Experts no independent evidence
    purpose: Select and momentum-merge layer-wise key shared experts so related tasks can transfer knowledge without overwriting key knowledge.
    Core algorithmic novelty; evidence is internal ablations and SOTA comparisons, not an external falsifiable prediction.

pith-pipeline@v1.1.0-grok45 · 31852 in / 3138 out tokens · 28556 ms · 2026-07-12T06:01:03.741193+00:00 · methodology

0 comments
read the original abstract

Multimodal Large Language Models (MLLMs) excel in diverse vision tasks, but full-parameter retraining is computationally expensive as real-world knowledge evolves. Existing continual learning methods often suffer from semantic entanglement in parameter spaces across tasks, impeding the continuous deployment of models. This challenge is especially pronounced in Anomaly Detection (AD), which exhibits triple heterogeneity across modalities, domains, and defect scale variability, significantly complicating multi-task knowledge transfer. In this paper, we propose CL-Anomaly, a parameter-efficient fine-tuning framework based on an isolation-sharing collaboration to enable continual learning for anomaly detection with MLLMs. We introduce the task-private expert PrivLoRA, which physically isolates task-specific subspaces in the parameter space to prevent semantic entanglement of anomaly knowledge in diverse scenarios. The Layer-Adaptive Shared Experts maintain cross-task representations within a unified feature space, enabling knowledge sharing between previous and new tasks. Furthermore, we propose a Layer-Adaptive Knowledge Transfer strategy that automatically selects and dynamically updates the layer-wise key shared experts of each task via a momentum-based mechanism, promoting effective knowledge transfer across related anomaly detection tasks. Extensive experiments across three continual learning scenarios for anomaly detection, including class-incremental, cross-domain, and cross-modal, demonstrate that CL-Anomaly outperforms state-of-the-art methods. Code is available at https://github.com/WenDongyp/CL-Anomaly.

Figures

Figures reproduced from arXiv: 2607.02930 by Ben Li, Caifeng Shan, Fang Zhao, Guo-Sen Xie, Kai Sun, Shuangqing Zhang, Wen Dong, Zhao Wang.

Figure 1
Figure 1. Figure 1: (Left) Parameter isolation methods prohibit cross-task knowledge transfer, pre￾venting knowledge complementarity. (Middle) Parameter sharing methods often lead to semantic entanglement, where parameter spaces of multiple tasks overlap and conflict. (Right) Our isolation–sharing collaboration framework maintains task-specific knowl￾edge within isolated parameter spaces while enabling cross-task knowledge tr… view at source ↗
Figure 2
Figure 2. Figure 2: Heatmaps of expert activation across different layers in the LoRA-based MoE architecture on WebAD and BMAD datasets. Layer-wise experts exhibit varying ac￾tivation preferences across tasks, which is widely observed across MLLMs of diverse scales and architectures, highlighting the importance of differentiating layer-wise ex￾perts based on their task preferences. HiDe-LLaVA [21] isolate parameter spaces for… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our method. (Left) CL-Anomaly consists of parallel task-private and task-shared paths. PrivLoRA is updated only within its task, extracting visual anchors with a multimodal encoder and assigning expert weights by similarity to task anchors. The task-shared path, termed Layer-Adaptive Shared Experts, dynamically selects experts with a gating function and updates parameters using a layer-adaptive… view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of CL-Anomaly, and the SOTA continual learning approaches under class-incremental learning on prac￾tical single-anomaly scenarios. 1 2 3 4 5 6 Training Stage 91.0 91.5 92.0 92.5 93.0 Accuracy (%) WebAD-1 WebAD-2 WebAD-3 WebAD-4 WebAD-5 WebAD-6 (a) 1 2 3 4 5 6 Task ID 80 82 84 86 88 90 92 94 Accuracy (%) Zero-shot Train on WebAD-1 Last (b) [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Average last accuracy of λ. 0.5 0.55 0.65 0.75 0.85 0.95 1.0 Parameter 84.0 84.5 85.0 85.5 86.0 Accuracy (%) [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Average last accuracy of β [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Training efficiency comparison of our method and representative approaches. Yes, there is a noticeable defect on the screw: a circular indentation or scratch on the neck area, which disrupts the smooth surface. No, there are no defects for the object in the image. The screw is well-constructed with a smooth surface and consistent threading, indicating it is in good condition. Are there any defects for the … view at source ↗
Figure 10
Figure 10. Figure 10: (Left) Many existing methods cannot dynamically update knowledge, often requiring the deployment of multiple specialized models to address evolving detection scenarios. (Right) Our method enables a single model to continually learn both intra￾task and cross-task anomaly knowledge, accommodating broader application scenarios. In addition, detection targets themselves often evolve. For instance, in manu￾fac… view at source ↗
Figure 11
Figure 11. Figure 11: Anomaly reasoning question example. prompt = """ You are an objective judge. Rate the model’s response eference compared to the reference answer on a scale from 0 to 10, considering semantic correctness, logical coherence, reasoning soundness, and completeness: - 10: perfectly correct, clear, and complete - 0: completely irrelevant or wrong Reference answer: {ans} Model response: {resp} Only output a sing… view at source ↗
Figure 12
Figure 12. Figure 12: Prompt for LLM-Score Evaluation. enabling a more holistic evaluation of reasoning results. In this study, we em￾ploy Qwen-3-Max as the judging model for LLM-Score, with the detailed scoring prompt provided in [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Comparison with CL-MoE in CIL across WebAD datasets. Each subfigure corresponds to one dataset [PITH_FULL_IMAGE:figures/full_fig_p028_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Analysis of PrivLoRA Routing Selection. (a) Heatmap of PrivLoRA visual￾anchor routing, with darker colors indicating higher expert weighting for the task. (b) Last accuracy improvement with manually corrected PrivLoRA routing. E Limitation During inference, each PrivLoRA is assigned a weight based on the cosine similar￾ity between an input’s visual anchor and stored task anchors, such that visual￾anchor r… view at source ↗
Figure 15
Figure 15. Figure 15: Visualization examples across datasets (Mvtec-AD, BMAD, Real3D-AD). Each subfigure corresponds to one object category [PITH_FULL_IMAGE:figures/full_fig_p031_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Anomaly reasoning examples across datasets (part 1) [PITH_FULL_IMAGE:figures/full_fig_p032_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Anomaly reasoning examples across datasets (part 2) [PITH_FULL_IMAGE:figures/full_fig_p033_17.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 18 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  2. [2]

    arXiv preprint arXiv:2309.16609 (2023) 16 W

    Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023) 16 W. Dong et al

    Pith/arXiv arXiv 2023

  3. [3]

    5-vl technical report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al.: Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    Pith/arXiv arXiv 2025

  4. [4]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Bao,J.,Sun,H.,Deng,H.,He,Y.,Zhang,Z.,Li,X.:Bmad:Benchmarksformedical anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4042–4053 (2024)

    2024

  5. [5]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Bergmann, P., Fauser, M., Sattlegger, D., Steger, C.: Mvtec ad–a comprehen- sive real-world dataset for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9592–9600 (2019)

    2019

  6. [6]

    arXiv preprint arXiv:2112.09045 (2021)

    Bergmann, P., Jin, X., Sattlegger, D., Steger, C.: The mvtec 3d-ad dataset for un- supervised 3d anomaly detection and localization. arXiv preprint arXiv:2112.09045 (2021)

    Pith/arXiv arXiv 2021

  7. [7]

    Cao,T.,Zhu,J.,Pang,G.:Anomalydetectionunderdistributionshift.In:Proceed- ings of the IEEE/CVF International Conference on Computer Vision. pp. 6511– 6523 (2023)

    2023

  8. [8]

    In: Euro- pean Conference on Computer Vision

    Cao, Y., Zhang, J., Frittoli, L., Cheng, Y., Shen, W., Boracchi, G.: Adaclip: Adapt- ing clip with hybrid learnable prompts for zero-shot anomaly detection. In: Euro- pean Conference on Computer Vision. pp. 55–72. Springer (2024)

    2024

  9. [9]

    arXiv preprint arXiv:2504.11914 (2025)

    Chao, Y., Liu, J., Tang, J., Wu, G.: Anomalyr1: A grpo-based end-to-end mllm for industrial anomaly detection. arXiv preprint arXiv:2504.11914 (2025)

    Pith/arXiv arXiv 2025

  10. [10]

    Advances in Neural Information Processing Systems37, 57817–57840 (2024)

    Chen, C., Zhu, J., Luo, X., Shen, H.T., Song, J., Gao, L.: Coin: A benchmark of continual instruction tuning for multimodel large language models. Advances in Neural Information Processing Systems37, 57817–57840 (2024)

    2024

  11. [11]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chen, G., Shen, L., Shao, R., Deng, X., Nie, L.: Lion: Empowering multimodal large language model with dual-level visual knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26540– 26550 (2024)

    2024

  12. [12]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Zhong, M., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 24185–24198 (2024)

    2024

  13. [13]

    In: Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition

    Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without mem- orizing. In: Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition. pp. 5138–5146 (2019)

    2019

  14. [14]

    In: ECAI 2025, pp

    Dong, W., Chu, G., Pan, Z., Xie, G.S., Shan, C., Zhao, F.: Region-aware compo- sitional context prompting for zero-shot anomaly detection. In: ECAI 2025, pp. 82–89. IOS Press (2025)

    2025

  15. [15]

    Feng, W., Hao, C., Zhang, Y., Han, Y., Wang, H.: Mixture-of-loras: An efficient multitasktuningforlargelanguagemodels.arXivpreprintarXiv:2403.03432(2024)

    Pith/arXiv arXiv 2024

  16. [16]

    arXiv preprint arXiv:2501.01957 (2025)

    Fu, C., Lin, H., Wang, X., Zhang, Y.F., Shen, Y., Liu, X., Cao, H., Long, Z., Gao, H., Li, K., et al.: Vita-1.5: Towards gpt-4o level real-time vision and speech interaction. arXiv preprint arXiv:2501.01957 (2025)

    Pith/arXiv arXiv 2025

  17. [17]

    Gao,J.,Li,Y.,Cao,Z.,Li,W.:Interleaved-modalchain-of-thought.In:Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 19520–19529 (2025)

    2025

  18. [18]

    arXiv preprint arXiv:2304.15010 (2023) CL-Anomaly: Layer-Adaptive MoE with MLLM for CL in AD 17

    Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023) CL-Anomaly: Layer-Adaptive MoE with MLLM for CL in AD 17

    Pith/arXiv arXiv 2023

  19. [19]

    arXiv preprint arXiv:2506.11672 (2025)

    Ge, C., Wang, X., Zhang, Z., Chen, H., Fan, J., Huang, L., Xue, H., Zhu, W.: Dynamic mixture of curriculum lora experts for continual multimodal instruction tuning. arXiv preprint arXiv:2506.11672 (2025)

    Pith/arXiv arXiv 2025

  20. [20]

    Sensors22(9), 3166 (2022)

    Givnan, S., Chalmers, C., Fergus, P., Ortega-Martorell, S., Whalley, T.: Anomaly detection using autoencoder reconstruction upon industrial motors. Sensors22(9), 3166 (2022)

    2022

  21. [21]

    arXiv preprint arXiv:2503.12941 (2025)

    Guo, H., Zeng, F., Xiang, Z., Zhu, F., Wang, D.H., Zhang, X.Y., Liu, C.L.: Hide- llava: Hierarchical decoupling for continual instruction tuning of multimodal large language model. arXiv preprint arXiv:2503.12941 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

    Advances in Neural Information Processing Systems37, 71162– 71187 (2024)

    He, H., Bai, Y., Zhang, J., He, Q., Chen, H., Gan, Z., Wang, C., Li, X., Tian, G., Xie, L.: Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems37, 71162– 71187 (2024)

    2024

  23. [23]

    In: International conference on machine learning

    Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Ges- mundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International conference on machine learning. pp. 2790–2799. PMLR (2019)

    2019

  24. [24]

    ICLR1(2), 3 (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. ICLR1(2), 3 (2022)

    2022

  25. [25]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Huai, T., Zhou, J., Wu, X., Chen, Q., Bai, Q., Zhou, Z., He, L.: Cl-moe: Enhanc- ing multimodal large language model with dual momentum mixture-of-experts for continual visual question answering. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 19608–19617 (2025)

    2025

  26. [26]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Huang, C., Jiang, A., Feng, J., Zhang, Y., Wang, X., Wang, Y.: Adapting visual- language models for generalizable anomaly detection in medical images. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 11375–11385 (2024)

    2024

  27. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Jeong, J., Zou, Y., Kim, T., Zhang, D., Ravichandran, A., Dabeer, O.: Winclip: Zero-/few-shot anomaly classification and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19606– 19616 (2023)

    2023

  28. [28]

    Proceedings of the national academy of sciences114(13), 3521–3526 (2017)

    Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences114(13), 3521–3526 (2017)

    2017

  29. [29]

    arXiv preprint arXiv:2104.08691 (2021)

    Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)

    Pith/arXiv arXiv 2021

  30. [30]

    arXiv preprint arXiv:2408.03326 (2024)

    Li, B., Zhang, Y., Guo, D., Zhang, R., Li, F., Zhang, H., Zhang, K., Zhang, P., Li, Y., Liu, Z., et al.: Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326 (2024)

    Pith/arXiv arXiv 2024

  31. [31]

    In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition

    Li, W., Xu, X., Gu, Y., Zheng, B., Gao, S., Wu, Y.: Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self- supervised learning network. In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition. pp. 22207–22216 (2024)

    2024

  32. [32]

    arXiv preprint arXiv:2101.00190 (2021)

    Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)

    Pith/arXiv arXiv 2021

  33. [33]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li, X., Zhang, Z., Tan, X., Chen, C., Qu, Y., Xie, Y., Ma, L.: Promptad: Learning prompts with only normal samples for few-shot anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16838–16848 (2024) 18 W. Dong et al

    2024

  34. [34]

    arXiv preprint arXiv:2508.09178 (2025)

    Li, Y., Cao, Y., Liu, C., Xiong, Y., Dong, X., Huang, C.: Iad-r1: Reinforcing con- sistent reasoning in industrial anomaly detection. arXiv preprint arXiv:2508.09178 (2025)

    Pith/arXiv arXiv 2025

  35. [35]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Li, Y., Cao, Y., Liu, C., Xiong, Y., Dong, X., Huang, C.: Iad-r1: Reinforcing con- sistent reasoning in industrial anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 40, pp. 6583–6591 (2026)

    2026

  36. [36]

    IEEE transactions on pattern anal- ysis and machine intelligence40(12), 2935–2947 (2017)

    Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern anal- ysis and machine intelligence40(12), 2935–2947 (2017)

    2017

  37. [37]

    In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

    Lin, B., Ye, Y., Zhu, B., Cui, J., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. pp. 5971–5984 (2024)

    2024

  38. [38]

    Applied Intelligence53(20), 23628–23640 (2023)

    Lin, J., He, Y., Xu, W., Guan, J., Zhang, J., Zhou, S.: Latent feature reconstruc- tion for unsupervised anomaly detection. Applied Intelligence53(20), 23628–23640 (2023)

    2023

  39. [39]

    Advances in Neural Information Processing Systems35, 16165–16177 (2022)

    Lin, S., Yang, L., Fan, D., Zhang, J.: Beyond not-forgetting: Continual learning with backward knowledge transfer. Advances in Neural Information Processing Systems35, 16165–16177 (2022)

    2022

  40. [40]

    Sensors23(21), 8750 (2023)

    Liu, F., Zhu, X., Feng, P., Zeng, L.: Anomaly detection via progressive reconstruc- tion and hierarchical feature fusion. Sensors23(21), 8750 (2023)

    2023

  41. [41]

    Advances in neural information processing systems36, 34892–34916 (2023)

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems36, 34892–34916 (2023)

    2023

  42. [42]

    In: Proceedings of the AAAI conference on artificial intelligence

    Liu, J., Wu, K., Nie, Q., Chen, Y., Gao, B.B., Liu, Y., Wang, J., Wang, C., Zheng, F.: Unsupervised continual anomaly detection with contrastively-learned prompt. In: Proceedings of the AAAI conference on artificial intelligence. vol. 38, pp. 3639– 3647 (2024)

    2024

  43. [43]

    Advances in Neural Information Processing Systems36, 30402–30415 (2023)

    Liu, J., Xie, G., Chen, R., Li, X., Wang, J., Liu, Y., Wang, C., Zheng, F.: Real3d- ad: A dataset of point cloud anomaly detection. Advances in Neural Information Processing Systems36, 30402–30415 (2023)

    2023

  44. [44]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, Z., Zhou, Y., Xu, Y., Wang, Z.: Simplenet: A simple network for image anomaly detection and localization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20402–20411 (2023)

    2023

  45. [45]

    Advances in neural information processing systems30(2017)

    Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. Advances in neural information processing systems30(2017)

    2017

  46. [46]

    arXiv preprint arXiv:1711.05101 (2017)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    Pith/arXiv arXiv 2017

  47. [47]

    Piao, H., Wu, Y., Wu, D., Wei, Y.: Federated continual learning via prompt-based dualknowledgetransfer.In:Forty-firstInternationalConferenceonMachineLearn- ing (2024)

    2024

  48. [48]

    Advances in Neural Information Processing Systems37, 97476–97498 (2024)

    Qiao, F., Mahdavi, M.: Learn more, but bother less: parameter efficient continual learning. Advances in Neural Information Processing Systems37, 97476–97498 (2024)

    2024

  49. [49]

    Advances in neural information processing systems32(2019)

    Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., Wayne, G.: Experience replay for continual learning. Advances in neural information processing systems32(2019)

    2019

  50. [50]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T., Gehler, P.: Towards total recall in industrial anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318–14328 (2022)

    2022

  51. [51]

    In: European conference on computer vision

    Tang, J., Lu, H., Xu, X., Wu, R., Hu, S., Zhang, T., Cheng, T.W., Ge, M., Chen, Y.C., Tsung, F.: An incremental unified framework for small defect inspection. In: European conference on computer vision. pp. 307–324. Springer (2024) CL-Anomaly: Layer-Adaptive MoE with MLLM for CL in AD 19

    2024

  52. [52]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Tao, F., Xie, G.S., Zhao, F., Shu, X.: Kernel-aware graph prompt learning for few-shot anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 7347–7355 (2025)

    2025

  53. [53]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Tien, T.D., Nguyen, A.T., Tran, N.H., Huy, T.D., Duong, S., Nguyen, C.D.T., Truong, S.Q.: Revisiting reverse distillation for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 24511–24520 (2023)

    2023

  54. [54]

    arXiv preprint arXiv:2302.13971 (2023)

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    Pith/arXiv arXiv 2023

  55. [55]

    arXiv preprint arXiv:2103.04257 (2021)

    Wang, G., Han, S., Ding, E., Huang, D.: Student-teacher feature pyramid matching for anomaly detection. arXiv preprint arXiv:2103.04257 (2021)

    Pith/arXiv arXiv 2021

  56. [56]

    Visual Intelligence4(1), 7 (2026)

    Wang, H., Yan, W., Lai, Q., Geng, X.: Temporal consistency-aware text-to-motion generation. Visual Intelligence4(1), 7 (2026)

    2026

  57. [57]

    In: Findings of the Association for Computational Linguistics: EMNLP 2023

    Wang, X., Chen, T., Ge, Q., Xia, H., Bao, R., Zheng, R., Zhang, Q., Gui, T., Huang, X.J.: Orthogonal subspace learning for language model continual learning. In: Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 10658–10671 (2023)

    2023

  58. [58]

    In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision

    Wang, Z., Che, C., Wang, Q., Li, Y., Shi, Z., Wang, M.: Smolora: Exploring and defying dual catastrophic forgetting in continual visual instruction tuning. In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision. pp. 177– 186 (2025)

    2025

  59. [59]

    In: Medical Imaging 2018: Computer- Aided Diagnosis

    Wei, Q., Ren, Y., Hou, R., Shi, B., Lo, J.Y., Carin, L.: Anomaly detection for medi- cal images based on a one-class classification. In: Medical Imaging 2018: Computer- Aided Diagnosis. vol. 10575, pp. 375–380. SPIE (2018)

    2018

  60. [60]

    Advances in Neural Information Pro- cessing Systems37, 69925–69975 (2024)

    Wu, J., Zhong, M., Xing, S., Lai, Z., Liu, Z., Chen, Z., Wang, W., Zhu, X., Lu, L., Lu, T., et al.: Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. Advances in Neural Information Pro- cessing Systems37, 69925–69975 (2024)

    2024

  61. [61]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Xu, J., Lo, S.Y., Safaei, B., Patel, V.M., Dwivedi, I.: Towards zero-shot anomaly detection and reasoning with multimodal large language models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 20370–20382 (2025)

    2025

  62. [62]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Zhang, X., Xu, M., Qiu, D., Yan, R., Lang, N., Zhou, X.: Mediclip: Adapting clip for few-shot medical image anomaly detection. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 458–468. Springer (2024)

    2024

  63. [63]

    IEEE Transactions on Medical Imaging40(12), 3641–3651 (2021)

    Zhao, H., Li, Y., He, N., Ma, K., Fang, L., Li, H., Zheng, Y.: Anomaly detection for medical images using self-supervised and translation-consistent features. IEEE Transactions on Medical Imaging40(12), 3641–3651 (2021)

    2021

  64. [64]

    arXiv preprint arXiv:2506.05453 (2025)

    Zhao, H., Zhu, F., Guo, H., Wang, M., Wang, R., Meng, G., Zhang, Z.: Mllm- cl: Continual learning for multimodal large language models. arXiv preprint arXiv:2506.05453 (2025)

    arXiv 2025

  65. [65]

    arXiv preprint arXiv:2411.12591 (2024)

    Zheng, H., Xu, T., Sun, H., Pu, S., Chen, R., Sun, L.: Thinking before look- ing: Improving multimodal llm reasoning via mitigating visual hallucination. arXiv preprint arXiv:2411.12591 (2024)

    Pith/arXiv arXiv 2024

  66. [66]

    arXiv preprint arXiv:2310.18961 (2023)

    Zhou, Q., Pang, G., Tian, Y., He, S., Chen, J.: Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961 (2023)

    arXiv 2023

  67. [67]

    In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition

    Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition. pp. 5871–5880 (2021) 20 W. Dong et al. Appendix for CL-Anomaly: Layer-Adaptive Mixture-of-Experts with Multimodal Large Language Model for Continu...

    arXiv 2021