Pith. sign in

REVIEW 4 major objections 4 minor 67 references

DAER claims that domain-isolated experts routed by distribution similarity let continually adapting video-AI models beat prior methods while nearly eliminating forgetting.

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 · deepseek-v4-flash

2026-08-01 14:33 UTC pith:ABVPT6UI

load-bearing objection Good engineering, useful benchmark, but the reported evaluation uses a different routing kernel than the method claims—so the MMD story is currently unsupported. the 4 major comments →

arxiv 2607.18716 v1 pith:ABVPT6UI submitted 2026-07-21 cs.CV

Continual Video-MLLM Adaptation over Evolving Domains

Rui Cheng , Meixing Shi , Yuxiang Cai , Jingcai Guo , Jianwei Yin , Zhi Chen This is my paper
classification cs.CV
keywords continual learningvideo question answeringVideo-MLLMmixture-of-expertsdomain-incremental learningdistribution-aware routingmaximum mean discrepancyparameter-efficient adaptation
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 proposes a way for a video-understanding language model to keep learning from a sequence of very different video domains — daily clips, traffic scenes, cartoons, abstract reasoning tests — without overwriting what it already knows. Its claim is that the standard fix of sharing one set of adapters across all domains causes cross-domain interference, so instead it isolates each domain's knowledge in its own small group of LoRA-style experts while keeping the pretrained model frozen. A non-parametric routing rule based on maximum mean discrepancy decides which experts to use, and a second routing stage guesses the input's domain at test time when no domain label is provided. On a ten-dataset video-QA benchmark, the authors report the approach beats prior continual-learning baselines by over three accuracy points and turns backward forgetting negative, meaning later training can nudge earlier-domain performance up rather than down.

Core claim

The central claim is that combining strict parameter isolation (a fresh pool of lightweight experts per domain) with distribution-aware routing yields a rehearsal-free, task-id-free solution to domain-incremental Video-MLLM adaptation. The authors argue that MMD between the current input's embedding and expert-level prototype reservoirs is a better routing signal than learned parametric gates, and that matching against domain prototypes in a learned discriminative subspace makes task identification reliable even when raw backbone features overlap. Empirically, DAER raises average VidQA accuracy from 64.38% to 67.59% over the strongest prior baseline on one large backbone while reducing backw

What carries the argument

The core object is the DAER adapter: a mixture-of-experts module inserted in parallel with the feed-forward network of each transformer layer, where each expert is a low-rank (LoRA) residual and each domain owns an isolated expert pool. Routing is driven by two non-parametric comparators: intra-domain routing scores an input by mean kernel similarity to per-expert prototype reservoirs (MMD-based), and inter-domain routing scores each candidate domain by kernel affinity in a discriminative subspace built by covariance regularization plus supervised inter-class projection. Adaptive domain merging averages expert weights when prototype-level MMD between a new and an old domain falls below a thr

Load-bearing premise

The load-bearing premise is that the training stream comes as a sequence of known, discrete tasks, one dataset at a time, so a fresh expert pool can be allocated at each switch; if real deployments present an unsegmented, continuously shifting stream, DAER has no mechanism for detecting when the domain has changed.

What would settle it

Give the trained system a stream whose domain shifts are not announced in advance and are much smaller/seamless than the ten benchmark datasets (e.g., smooth weather or style drifts), and measure routing accuracy and accuracy decay: if routing accuracy falls toward chance or forgetting rises sharply, the central claim about task-free inference is unsupported.

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

If this is right

  • Continually adapted video QA models can keep improving across ten heterogeneous domains, with average accuracy ~67.6% and backward forgetting −0.14 on InternVideo2.5-8B; negative BWF implies later domains help earlier ones.
  • No domain label is needed at test time: the inter-domain router predicts the task, and its accuracy closely tracks the oracle-task-id performance (gap 0.54 points versus 2.64 for the strongest baseline).
  • Parameter growth is bounded by adaptive merging of similar domains; the added inference cost of matching against domain prototypes stays under a few percent of total latency even at an estimated 100-domain scale.
  • The approach transfers to a second Video-MLLM (VideoLLaMA3) and to 20-task streams, and also retains a separately learned tracking capability when the task stream includes a non-QA video task.

Where Pith is reading between the lines

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

  • The training pipeline still assumes known, segmented tasks; a natural next step is an online variant that detects domain shifts and decides when to spawn, split, or merge experts without a pre-segmented task stream.
  • Because both routing stages depend on the quality of the learned domain and expert prototype reservoirs, the approach may be sensitive to label noise and to how the warm-up data are sampled; both are testable by ablating reservoir size and warm-up data.
  • The 'negative forgetting' result suggests that prototype-driven routing may let future tasks act as a regularizer for older experts, a phenomenon worth testing on longer, harder domain chains where an oracle task boundary is not available.

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

4 major / 4 minor

Summary. The paper proposes DAER, a continual-learning framework for Video-MLLMs that adapts to sequentially arriving domains. It freezes the pretrained backbone and maintains per-domain LoRA-based expert pools, with an intra-domain MMD/prototype-matching router for expert selection, an inter-domain router for task identification at inference, adaptive domain merging, and a warm-up optimization stage. The method is evaluated on a newly curated benchmark of ten VidQA datasets using InternVideo2.5-8B and VideoLLaMA3, reporting state-of-the-art average accuracy and improved backward forgetting over prior continual-learning baselines.

Significance. If the reported results are robust, this is a useful contribution to continual Video-MLLM adaptation. The paper builds a compact but challenging domain-incremental VidQA benchmark, compares against a range of existing CL methods, and includes extensive ablations, efficiency analyses, and an oracle-task-id comparison. The release of code is also a positive. The core idea of combining strict parameter isolation with distribution-aware routing is timely and, provided the implementation/description gap is fixed, of interest to both the continual-learning and multimodal LLM communities.

major comments (4)
  1. [§3.2, Eq. (4) and Appendix E, Table 6] There is an internal inconsistency between the described routing mechanism and the evaluated system. The main text states that K in Eq. (4) is a multi-kernel Gaussian (RBF) function and derives the routing signal from MMD. However, Appendix E reports that the best configuration is 'Cosine (Ours)' (67.59%) while the multi-kernel Gaussian reaches 67.48%. Since the headline numbers in Tables 1 and 2 are produced by the cosine-based system, the SOTA claim is not tied to the MMD mechanism described in the paper. Please either align the description with the implementation or rerun the main experiments with the described multi-kernel Gaussian and report those results as primary.
  2. [Tables 1 and 2] All results appear to come from a single run, with no error bars or significance tests. The average gain over DAM is 3.21 points on InternVideo2.5 but only 1.82 points on VideoLLaMA3, and per-domain differences are sometimes negative. Given the small margins on some domains, run-to-run variation could affect the conclusions. The authors should report mean and standard deviation over at least three seeds, and should state how the final hyperparameters were selected (validation vs. test).
  3. [Fig. 5 and Tables 1–2] The sensitivity study appears to select the final configuration (n_e=8, tau=0.006, 64 expert prototypes) based on performance on the same benchmark that is then used for the main results. This is a form of selection bias that can make reported accuracies optimistic. The paper should either use a held-out validation split for hyperparameter selection or show that the reported result is stable across the range of hyperparameters rather than only at the chosen point.
  4. [§3.1–§3.2] DAER assumes that the training stream is segmented into known tasks: Section 3.1 states that 'each task t is associated with a dataset D_t', and Section 3.2 allocates a dedicated MoE_t per task. The method removes task identity only at inference; it does not detect when a domain shift occurs during training. This is a significant scope limitation relative to the title's 'evolving domains' and the abstract's 'data arrives continuously'. The authors should explicitly state this assumption and discuss how the framework could be extended to unknown shift points.
minor comments (4)
  1. [Appendix A] Appendix A describes the Top-k routing strategy as 'Top-k = routing strategy' without giving the numeric k value used in Eq. (5). Please specify the actual k employed in the experiments.
  2. [Table 2] For VideoLLaMA3, only DAM is compared as a prior method. To support the claim of consistent improvement over prior art, include at least one strong prompt-based or adapter-based baseline (e.g., S-Prompts or MoE-Adapters) for this backbone as well.
  3. [Figure 3] The 20-task curves are hard to read because the figure is small and the task names are abbreviated. Provide a table of numeric results or a larger figure.
  4. [Table 6, Kernel row] The label 'Cosine (Ours)' conflicts with the main-text specification of the kernel in Eq. (4). If cosine was the intended kernel, the main text should be corrected; otherwise the table should be clarified.

Circularity Check

0 steps flagged

No circularity found: DAER's derivation is self-contained; flagged concerns are consistency/assumption issues, not circular reasoning.

full rationale

Walking the claimed derivation chain: the intra-domain routing score (Eq. 4) is defined as a kernel similarity to expert prototypes and used to select experts; this is a design mechanism, not a quantity fitted to the benchmark and then re-reported as a prediction. The inter-domain task predictor (Eqs. 6-8) is trained with task-level prototype information and evaluated on unseen test samples; no test-set accuracy or forgetting value is used to set its parameters or constants. Adaptive domain merging (Eq. 9) uses MMD between prototype reservoirs to decide whether to merge, again an algorithmic rule rather than a fitted result. Headline numbers in Tables 1-2 are comparisons against external baselines on public VidQA datasets, so the central claim is not equivalent to an input by construction. No load-bearing self-citation chain exists: references such as DAM [11] and collaborative prompting [5] are external prior work, and no author-specific uniqueness theorem is invoked. The manuscript nevertheless contains an inconsistency worth noting as a correctness risk, not circularity: Section 3.2/Eq. (4) states that K is a multi-kernel Gaussian (RBF) function, while Appendix E/Table 6 reports the best configuration as Cosine with 67.59% (multi-Gaussian 67.48%). Also, the known-task-boundary training assumption in Sec. 3.1 ('Each task t is associated with a dataset D_t') limits the evolving-stream motivation, and Appendix A leaves 'Top-k = routing strategy' unspecified. These are implementation/assumption gaps, not cases where a result reduces to its inputs. Hyperparameters tuned on the benchmark are a selection-bias concern, not circularity. Hence score 0.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 3 invented entities

DAER is an empirical engineering contribution with no new fundamental theory. Its central claims rest on standard deep learning assumptions (frozen backbone + LoRA can adapt, prototypes represent distributions) plus an ad hoc merge threshold. The largest hidden cost is the assumption that task boundaries are visible during training.

free parameters (8)
  • Number of experts per domain (n_e) = 8
    Selected from sensitivity study (Fig. 5a) on the benchmark; no validation protocol is described.
  • LoRA rank = 32
    Chosen by hand; stated in Appendix A.
  • Top-k activated experts = not specified
    The appendix text says 'a Top-k = routing strategy is employed' but the k value is omitted.
  • ADM merging threshold τ = 0.006
    Tuned via sensitivity study (Fig. 5b); merging is triggered when MMD between domain reservoirs is below τ.
  • Expert prototype count n_p = 64
    Best value in Fig. 5c; number of prototypes maintained per expert.
  • Domain prototype count n_m = 128
    Appendix H states each domain reservoir stores 128 prototypes; sensitivity shown in Fig. 5d.
  • Prototype mixing ratio α in Eq. (6) = not stated
    Ratio for combining text and video features into domain prototypes; the value is not reported.
  • Warm-up data fraction = 30%
    Fraction of current task training data used in the warm-up stage; chosen by hand.
axioms (5)
  • domain assumption The training stream is pre-segmented into discrete datasets, and the model knows when a new task begins during training; only inference is task-free.
    Section 3.1 defines per-task datasets D_t and states 'We can only access D_t when training on task t'; Section 3.2 allocates MoE_t per task. In a truly continuous stream this boundary knowledge is unavailable.
  • domain assumption MMD/cross-kernel routing score is a valid proxy for which expert will produce the most accurate output.
    Eq. (4) assumes closer distribution match implies better expert fit; no theoretical or independent empirical validation of this proxy is given.
  • domain assumption Prototype reservoirs are representative of the underlying domain distributions.
    Warm-up and reservoir sampling are expected to make prototypes representative, but this is not guaranteed under intra-domain shift or small sample sizes.
  • ad hoc to paper Merging similar domains when MMD < τ does not harm final performance.
    Table 6 shows ADM slightly reduces average accuracy compared to no merging (67.59 vs 67.68); the threshold τ is an ad hoc heuristic.
  • domain assumption A frozen Video-MLLM backbone plus LoRA-style experts can capture domain-specific knowledge without cross-domain interference.
    This is the central design premise of the paper; the experiments support it for the tested benchmark but it is not proven generally.
invented entities (3)
  • Expert-level prototype reservoir P_i_e no independent evidence
    purpose: Stores per-expert prototype embeddings used for MMD-based intra-domain routing.
    Internal data structure; no external falsifiable handle beyond downstream accuracy.
  • Domain-level prototype reservoir M_t no independent evidence
    purpose: Stores per-task prototypes for inter-domain task prediction.
    Internal memory; only evaluated indirectly through routing accuracy and downstream task accuracy.
  • Discriminative subspace projection W = W_reg W_disc no independent evidence
    purpose: Projects features into a subspace that increases inter-domain margins for task identification.
    Learned with task labels; no evidence outside the paper's benchmark.

pith-pipeline@v1.3.0-alltime-deepseek · 22030 in / 15524 out tokens · 142015 ms · 2026-08-01T14:33:23.257543+00:00 · methodology

0 comments
read the original abstract

Video multimodal large language models have shown strong capability in video understanding, yet their adaptation to sequentially evolving domains remains underexplored. In real-world deployments, video data often arrives continuously from heterogeneous domains, requiring the model to acquire new domain-specific knowledge without overwriting previously learned capabilities. Existing continual learning methods typically rely on shared adaptation spaces, which can induce severe cross-domain interference and catastrophic forgetting. We propose Distribution-Aware Expert Routing, a parameter-efficient framework for continual Video-MLLM adaptation over evolving domains. DAER maintains domain-isolated lightweight experts while keeping the pretrained Video-MLLM backbone frozen, thereby decoupling domain-specific adaptation from the general multimodal knowledge of the pretrained model. To enable fine-grained specialization, we introduce an intra-domain distribution-aware routing mechanism that matches each input to expert-level prototype reservoirs using MMD. To address the absence of task identities at inference time, we further propose an inter-domain routing mechanism that performs prototype matching in a discriminative subspace for robust domain identification. In addition, we introduce adaptive domain merging to improve parameter scalability and adopt a two-stage optimization strategy to stabilize expert specialization during continual learning. We evaluate DAER by curating a domain-incremental benchmark built from ten VidQA datasets covering diverse visual environments and reasoning demands. Experiments on two strong Video-MLLM backbones show that DAER consistently outperforms prior methods.

Figures

Figures reproduced from arXiv: 2607.18716 by Jianwei Yin, Jingcai Guo, Meixing Shi, Rui Cheng, Yuxiang Cai, Zhi Chen.

Figure 1
Figure 1. Figure 1: Overview of DAER for continual Video-MLLM adap [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the DAER framework for continual Video-MLLM adaptation over evolving domains. DAER performs [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Results under the 20-task continual learning setting. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on four important components. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity study on four critical hyper-parameters. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of domain routing features. [PITH_FULL_IMAGE:figures/full_fig_p008_6.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 · 11 linked inside Pith

  1. [1]

    Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cuc- chiara, and Babak Ehteshami Bejnordi. 2020. Conditional channel gated networks for task-aware continual learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 3931–3940

    2020

  2. [2]

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al

  3. [3]

    Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. 2017. Expert gate: Lifelong learning with a network of experts. InProceedings of the IEEE conference on computer vision and pattern recognition. 3366–3375

    2017

  4. [4]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners.Advances in neural information processing systems33 (2020), 1877–1901

    2020

  5. [5]

    Chen Cai, Zheng Wang, Jianjun Gao, Wenyang Liu, Ye Lu, Runzhong Zhang, and Kim-Hui Yap. 2024. Empowering large language model for continual video question answering with collaborative prompting. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 3921–3932

    2024

  6. [6]

    Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. 2021. Co2l: Contrastive continual learning. InProceedings of the IEEE/CVF International conference on computer vision. 9516–9525

    2021

  7. [7]

    Bin Chen, Yuxiang Cai, Yadan Luo, Yi Zhang, Jianwei Yin, and Zhi Chen*. 2026. Spectral Evolution-Guided Token Pruning in Multimodal Large Language Models. InECCV 2026

    2026

  8. [8]

    Zengjie Chen, Yuxiang Cai, Jingcai Guo, Taotao Cai, Jianwei Yin, and Zhi Chen*

  9. [9]

    Zhi Chen, Xin Yu, Xiaohui Tao, Yan Li, and Zi Huang. 2025. Cluster-aware prompt ensemble learning for few-shot vision–language model adaptation.Pattern Recognition(2025), 112596

    2025

  10. [10]

    Feng Cheng, Xizi Wang, Jie Lei, David Crandall, Mohit Bansal, and Gedas Berta- sius. 2023. Vindlu: A recipe for effective video-and-language pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. 10739–10750

    2023

  11. [11]

    Feng Cheng, Ziyang Wang, Yi-Lin Sung, Yan-Bo Lin, Mohit Bansal, and Gedas Bertasius. 2024. Dam: Dynamic adapter merging for continual video qa learning. arXiv preprint arXiv:2403.08755(2024)

    Pith/arXiv arXiv 2024

  12. [12]

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Se- bastian Gehrmann, et al. 2023. Palm: Scaling language modeling with pathways. Journal of machine learning research24, 240 (2023), 1–113

    2023

  13. [13]

    Yi Dai, Hao Lang, Yinhe Zheng, Bowen Yu, Fei Huang, and Yongbin Li. 2023. Do- main incremental lifelong learning in an open world. InFindings of the Association for Computational Linguistics: ACL 2023. 5844–5865

    2023

  14. [14]

    Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. 2019. Lasot: A high-quality benchmark for large-scale single object tracking. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5374–5383

    2019

  15. [15]

    Raia Hadsell, Dushyant Rao, Andrei A Rusu, and Razvan Pascanu. 2020. Em- bracing change: Continual learning in deep neural networks.Trends in cognitive sciences24, 12 (2020), 1028–1040

    2020

  16. [16]

    Christian Henning, Maria Cervera, Francesco D’Angelo, Johannes Von Oswald, Regina Traber, Benjamin Ehret, Seijin Kobayashi, Benjamin F Grewe, and Joao Sacramento. 2021. Posterior meta-replay for continual learning.Advances in neural information processing systems34 (2021), 14135–14149

    2021

  17. [17]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models.Iclr1, 2 (2022), 3

    2022

  18. [18]

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. 1991. Adaptive mixtures of local experts.Neural computation3, 1 (1991), 79–87

    1991

  19. [19]

    Kyung-Min Kim, Min-Oh Heo, Seong-Ho Choi, and Byoung-Tak Zhang. 2017. Deepstory: Video story qa by deep embedded memory networks.arXiv preprint arXiv:1707.00836(2017)

    Pith/arXiv arXiv 2017

  20. [20]

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. 2017. Overcoming catastrophic forgetting in neural networks.Proceedings of the national academy of sciences114, 13 (2017), 3521– 3526

    2017

  21. [21]

    Christiaan Lamers, René Vidal, Nabil Belbachir, Niki Van Stein, Thomas Bäeck, and Paris Giampouras. 2023. Clustering-based domain-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 3384– 3392

    2023

  22. [22]

    Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. 2021. Less is more: Clipbert for video-and-language learning via sparse sampling. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 7331–7341

    2021

  23. [23]

    Hongbo Li, Sen Lin, Lingjie Duan, Yingbin Liang, and Ness B Shroff. 2024. Theory on mixture-of-experts in continual learning.arXiv preprint arXiv:2406.16437 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning. PMLR, 19730–19742

    2023

  25. [25]

    Zhizhong Li and Derek Hoiem. 2017. Learning without forgetting.IEEE transac- tions on pattern analysis and machine intelligence40, 12 (2017), 2935–2947

    2017

  26. [26]

    Haowei Lin, Yijia Shao, Weinan Qian, Ningxin Pan, Yiduo Guo, and Bing Liu

  27. [27]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning.Advances in neural information processing systems36 (2023), 34892–34916

    2023

  28. [28]

    David Lopez-Paz and Marc’Aurelio Ranzato. 2017. Gradient episodic memory for continual learning.Advances in neural information processing systems30 (2017)

    2017

  29. [29]

    Marc Masana, Xialei Liu, Bartłomiej Twardowski, Mikel Menta, Andrew D Bag- danov, and Joost Van De Weijer. 2022. Class-incremental learning: survey and performance evaluation on image classification.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 5 (2022), 5513–5533

    2022

  30. [30]

    James L McClelland, Bruce L McNaughton, and Randall C O’Reilly. 1995. Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory.Psychological review102, 3 (1995), 419

    1995

  31. [31]

    M Jehanzeb Mirza, Marc Masana, Horst Possegger, and Horst Bischof. 2022. An efficient domain-incremental learning approach to drive in all weather condi- tions. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 3001–3011

    2022

  32. [32]

    Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Mar- keeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. 2023. Perception test: A diagnostic benchmark for multimodal video models.Advances in Neural Information Processing Systems36 (2023), 42748–42761

    2023

  33. [33]

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

    2017

  34. [34]

    Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. 2018. Learning to learn without forgetting by maximizing transfer and minimizing interference.arXiv preprint arXiv:1810.11910(2018)

    Pith/arXiv arXiv 2018

  35. [35]

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538(2017)

    Pith/arXiv arXiv 2017

  36. [36]

    Haizhou Shi and Hao Wang. 2023. A unified approach to domain incremental learning with memory: Theory and algorithm.Advances in Neural Information Processing Systems36 (2023), 15027–15059

    2023

  37. [37]

    James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira

  38. [38]

    Xiaomeng Song, Yucheng Shi, Xin Chen, and Yahong Han. 2018. Explore multi- step reasoning in video question answering. InProceedings of the 26th ACM international conference on Multimedia. 239–247

    2018

  39. [39]

    Tianqi Tang, Shohreh Deldari, Hao Xue, Celso De Melo, and Flora Salim. 2024. Vilco-bench: Video language continual learning benchmark.Advances in Neural Information Processing Systems37 (2024), 70213–70229

    2024

  40. [40]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971(2023)

    Pith/arXiv arXiv 2023

  41. [41]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11909–11919

  42. [42]

    Andrés Villa, Juan León Alcázar, Motasem Alfarra, Kumail Alhamoud, Julio Hurtado, Fabian Caba Heilbron, Alvaro Soto, and Bernard Ghanem. 2023. Pivot: Prompting for video continual learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 24214–24223

    2023

  43. [43]

    Tianqi Wang, Jingcai Guo, Depeng Li, and Zhi Chen. 2025. On the discrimination and consistency for exemplar-free class incremental learning. InIJCAI 2025

    2025

  44. [44]

    Wanqi Wang, Jingcai Guo, Yuxiang Cai, and Zhi Chen*. 2026. Learning Multi- Modal Prototypes for Cross-Domain Few-Shot Object Detection. InCVPR 2026 Findings

    2026

  45. [45]

    Gido M Van de Ven, Tinne Tuytelaars, and Andreas S Tolias. 2022. Three types of incremental learning.Nature Machine Intelligence4, 12 (2022), 1185–1197

    2022

  46. [46]

    Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al . 2025. Internvideo2. 5: MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Rui Cheng et al. Empowering video mllms with long and rich context modeling.arXiv preprint arXiv:2501.12386(2025)

    Pith/arXiv arXiv 2025

  47. [47]

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. 2022. Dualprompt: Complementary prompting for rehearsal-free continual learning. InEuropean conference on computer vision. Springer, 631–648

    2022

  48. [48]

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. 2022. Learning to prompt for continual learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 139–149

    2022

  49. [49]

    Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. 2022. S-prompts learning with pre-trained transformers: An occam’s razor for domain incremental learning. Advances in Neural Information Processing Systems35 (2022), 5682–5695

    2022

  50. [50]

    Junbin Xiao, Pan Zhou, Tat-Seng Chua, and Shuicheng Yan. 2022. Video graph transformer for video question answering. InEuropean Conference on Computer Vision. Springer, 39–58

    2022

  51. [51]

    Binzhu Xie, Sicheng Zhang, Zitang Zhou, Bo Li, Yuanhan Zhang, Jack Hessel, Jingkang Yang, and Ziwei Liu. 2024. Funqa: Towards surprising video compre- hension. InEuropean Conference on Computer Vision. Springer, 39–57

    2024

  52. [52]

    Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. 2017. Video question answering via gradually refined atten- tion over appearance and motion. InProceedings of the 25th ACM international conference on Multimedia. 1645–1653

    2017

  53. [53]

    Tianqi Wei, Xin Yu, Zhi Chen, Scott Chapman, and Zi Huang. 2026. Benchmarking Vision-Language Models for Microscopic Plant Image Understanding. InECCV 2026

    2026

  54. [54]

    Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. 2022. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems35 (2022), 124–141

    2022

  55. [55]

    Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. 2024. Boosting continual learning of vision-language models via mixture-of- experts adapters. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 23219–23230

    2024

  56. [56]

    Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. 2024. Timesuite: Improving mllms for long video understanding via grounded tuning.arXiv preprint arXiv:2410.19702(2024)

    Pith/arXiv arXiv 2024

  57. [57]

    Li Xu, He Huang, and Jun Liu. 2021. Sutd-trafficqa: A question answering benchmark and an efficient network for video reasoning over traffic events. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9878–9888

    2021

  58. [58]

    Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. 2025. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106(2025)

    Pith/arXiv arXiv 2025

  59. [59]

    Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. 2024. Llava-video: Video instruction tuning with synthetic data.arXiv preprint arXiv:2410.02713(2024)

    Pith/arXiv arXiv 2024

  60. [60]

    Zecheng Zhao, Zhi Chen, Zi Huang, Shazia Sadiq, and Tong Chen. 2025. Continual text-to-video retrieval with frame fusion and task-aware routing. InSIGIR 2025. 1011–1021

    2025

  61. [61]

    Friedemann Zenke, Ben Poole, and Surya Ganguli. 2017. Continual learning through synaptic intelligence. InInternational conference on machine learning. Pmlr, 3987–3995

    2017

  62. [62]

    Zecheng Zhao, Selena Song, Tong Chen, Zhi Chen, Shazia Sadiq, and Yadan Luo. 2025. Are Synthetic Videos Useful? A Benchmark for Retrieval-Centric Evaluation of Synthetic Videos. InACM MM 2025

    2025

  63. [63]

    Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. 2024. Class-incremental learning: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence46, 12 (2024), 9851–9873. Continual Video-MLLM Adaptation over Evolving Domains MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Appendix A More Implementation Details...

    arXiv 2024

  64. [65]

    Zecheng Zhao, Zhi Chen, Zi Huang, Shazia Sadiq, and Tong Chen. 2026. Gener- ative Recall, Dense Reranking: Learning Multi-View Semantic IDs for Efficient Text-to-Video Retrieval.arXiv preprint arXiv:2601.21193(2026)

    arXiv 2026

  65. [2022]

    Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems35 (2022), 23716–23736

    2022

  66. [2023]

    Class incremental learning via likelihood ratio based task prediction.arXiv preprint arXiv:2309.15048(2023)

    Pith/arXiv arXiv 2023

  67. [2026]

    InECCV 2026

    Accelerating Multimodal Large Language Models with Prior-Corrected Token Reduction. InECCV 2026

    2026