Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Scalable Strategies for Continual Learning with Replay

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read This paper claims replay-based continual learning can match standard accuracy with up to 65% fewer replay samples by moving replay into a post-task consolidation phase and merging weights sequentially.

desk verdict The consolidation-phase replay framework is a genuinely useful idea, but the headline replay-savings numbers need cleaner evaluation hygiene (a locked test split, seed variance) before I'd take them at face value. read the letter →

arxiv 2505.12512 v1 pith:PNIY5VYX submitted 2025-05-18 cs.LG cs.AIcs.CV

Truman Hickok This is my paper · ORCID
classification cs.LGcs.AIcs.CV
keywords continuallearningreplayconsolidationlow-rankadaptationmodelmergingclass-incrementalpre-trainingsampleefficiency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to make replay-based continual learning much cheaper without giving up accuracy. Replay, showing a model stored examples of older tasks while it learns a new one, is the standard fix for catastrophic forgetting but roughly doubles training cost. The paper's central empirical claim is that replay can be redistributed: learn a task with a low replay ratio, then spend a dedicated post-task consolidation phase training on replayed samples, optionally with low-rank adaptation and with the model's weights merged before and after each task. In the reported experiments, consolidation alone matches a standard 1:1 replay baseline with up to 55% fewer replay samples, and the full combination matches it with up to 65% fewer samples. That is the paper's central quantitative claim, obtained on image-classification task sequences of up to twenty tasks.

What carries the argument

The load-bearing mechanism is a two-phase replay schedule: during task learning the replay ratio $RR = N_{\text{replay}}/N_{\text{task}}$ is lowered, and the samples thereby saved are spent in a post-task consolidation phase whose size is set by a consolidation step rate $CSR \in [0,1]$, with efficiency tracked by the total replay percentage (TRP). The second mechanism is the sequential merge $\theta_t = (1-\alpha)\theta_{t-1} + \alpha\theta^{*}_t$, a one-line interpolation between the pre-task and post-task checkpoints that gives EMA-like regularization without storing a shadow parameter after every gradient step. Low-rank adaptation (LoRA), a parameter-efficient fine-tuning method that learns only small low-rank update matrices per task and then merges them into the base weights, enters as a situational regularizer that is most useful when replay is scarce or tasks are small.

What would settle it

Re-run the best configuration (sequential merging plus consolidation at a 35% total replay percentage) with a strictly disjoint test set used only for final accuracy, while all checkpoint selection and hyperparameter tuning use only validation data; if the 35% run no longer matches the 1:1 replay baseline within one standard deviation across seeds, the central 65% replay-reduction claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the sample cost of replay can be separated from its accuracy benefit by splitting training into a task-learning phase with a reduced replay ratio and a consolidation phase that trains only on replayed samples from a balanced distribution over previous classes. The total replay percentage (TRP) measures replay samples actually used relative to what a standard 1:1 replay schedule would have used, and a consolidation step rate (CSR) controls what fraction of the samples saved by lowering the replay ratio is spent in consolidation. On top of this, sequential merging, which updates the model after each task by interpolating its pre- and post-task weights as $\theta_t = (1-\alpha)\theta_{t-1} + \alpha\theta^{*}_t$, is shown to match an exponential moving average while needing no per-gradient-step checkpoints. In the reported experiments the combined toolset, low-rank adaptation applied where needed plus consolidation plus sequential merging, matches the accuracy of the 1:1 replay baseline while using up to 65% fewer replay samples, and sequential merging becomes increasingly preferable to parallel merging as the number of tasks grows.

Load-bearing premise

The replay-reduction percentages assume the reported accuracies come from a held-out test set, but the paper never defines the train/test split and Section 9 says hyperparameters are chosen from validation accuracy, so if the same data are used for selection and reporting the savings could be inflated.

Editorial extensions

If this is right

  • The standard 1:1 replay ratio is not a cost floor: at the same total replay percentage, consolidation raises accuracy (for example, 76.1 versus 73.8 in class-incremental learning).
  • A model can match a full 1:1 replay baseline with about a third to half of the replay samples, which means fewer stored samples and shorter per-task training time.
  • Sequential merging offers a cheap alternative to an exponential moving average, with comparable accuracy and no per-step checkpoint storage, making it suitable for long task streams.
  • Parallel task-vector merging falls behind sequential merging as the number of tasks grows, so sequential merging is the better fit for continual learning.
  • Low-rank adaptation should be used selectively: it prevents collapse at low replay ratios and in continual pre-training with small tasks, while full fine-tuning remains better when replay is plentiful.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct testable extension is to replace the balanced consolidation distribution with one biased toward classes whose accuracy dropped during task learning; if per-class accuracy is measurable online, targeted consolidation should beat balanced consolidation at the same TRP.
  • The paper's scaling argument implies that as logits become better calibrated in larger models, the replay ratio during task learning could approach zero and consolidation could run offline in batches, extending the reported saving beyond 65%.
  • The same consolidation idea transfers to off-policy reinforcement learning, where replay ratios are already a central control knob: moving some replay updates into a dedicated consolidation phase may reduce total environment interactions for a fixed performance target.
  • Delaying consolidation until several tasks have accumulated should improve the trade-off further, since the replay distribution is then larger and more diverse; a clean published version of that experiment would sharpen the paper's scalability claim.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes and evaluates three techniques for making replay-based continual learning more sample-efficient: low-rank adaptation (LoRA), a phasic 'consolidation' strategy that redistributes saved replay samples into a post-task training phase, and 'sequential merging,' an adaptation of model merging that interpolates pre- and post-task weights after each task. The experiments compare these methods, individually and combined, against full fine-tuning and a 1:1 replay baseline across task-incremental, class-incremental, and continual pre-training settings. The central quantitative claim is that sequential merging plus consolidation reaches the same accuracy as the 1:1 baseline while using 65% fewer replay samples (Table 2), with consolidation alone saving up to 55% (Section 5.2). The paper also presents analyses of LoRA versus full fine-tuning under varying task sizes and replay ratios.

Significance. If the reported replay-reduction numbers are reliable, the paper would make a useful empirical contribution: it introduces a clean way to account for total replay usage (TRP), demonstrates that reallocating replay samples to a post-task phase can improve sample efficiency, and shows that sequential merging is a cheap and effective regularizer in continual learning. The main ideas are simple, intuitive, and plausibly transferable. The paper also makes good use of a controlled replay budget rather than treating replay as an unconstrained resource. However, the significance is currently limited by the lack of a clearly specified evaluation protocol: without a defined hold-out test split and with only three task-split seeds and no variance estimates, the headline equality in Table 2 (73.8 vs. 73.8) is not adequately supported.

major comments (3)
  1. [Section 9 / Section 3] The paper never defines a train/validation/test split. Section 9 states that the learning rate is swept and the best post-epoch checkpoint is selected 'based on validation accuracy,' but no test set is mentioned anywhere in the manuscript. If the reported accuracies include validation data used for checkpoint selection and hyperparameter tuning, then the comparison between configurations is biased in favor of the tuned methods, and the exact equality in Table 2 (Seq.+Consol. at TRP=35% vs. No merge/consol. at TRP=100%, both 73.8 in CIL) becomes uninterpretable. The authors must specify a fixed hold-out test split, describe how validation is used for model selection, and report test-set accuracies separately from validation accuracies.
  2. [Section 3] The statement that standard deviations are omitted because they represent variance across task splits, not variance across runs, is not a valid reason for omitting them: the task-split variance is precisely the variance that determines whether the observed differences between TRP=35% and TRP=100% rows are meaningful. With only 3 seeds and differences as small as 0.0–0.3 points (Table 2), the claim that Seq.+Consol. matches the baseline could easily flip with one additional seed. Report per-seed results, standard deviations, or confidence intervals for the headline rows in Table 2 and Figure 4.
  3. [Section 7] The paper states that for each algorithmic configuration 'we tune all available hyperparameters, which includes loss coefficients, merging alpha, consolidation step rates, replay ratios, and LoRA rank.' When combined with the absence of a locked test set, this is a multiple-comparisons procedure: selecting the best hyperparameters for each configuration on the same data that produces the reported accuracies can inflate the apparent advantage of the more flexible configurations. The authors should describe the hyperparameter selection protocol (e.g., a fixed validation split, a fixed search budget, early stopping on validation) and explicitly confirm that all reported numbers are measured on a test set that was never used for any selection decision.
minor comments (4)
  1. [Section 5.2] The text says 'Table 4 shows performances across consolidation step rates...' but there is no Table 4 in the manuscript; the data appear in Figure 4, which is a figure containing a table-like block. Re-number the reference or convert the block into a proper table.
  2. [Section 7] The sentence 'which is less than standalone consolidation requires for the same performance (see Table??)' contains an unresolved cross-reference. The referenced table should be identified (likely Table 2 or Figure 4).
  3. [Section 3] The experimental setup omits several details that are needed for reproducibility: the specific vision transformer variant (e.g., ViT-B/16 or ViT-L/14), input resolution, optimizer and batch size, and the total number of training steps per task. Please add these to the supplementary material.
  4. [Abstract and Section 5.2 vs. Section 7] The abstract mentions 'up to 55% less replay samples' while the introduction and conclusion mention 'up to 65% less'. These refer to different configurations (consolidation alone vs. consolidation plus sequential merging), but the distinction is not made explicit in the abstract. State the configuration that achieves the 65% reduction (Seq.+Consol. at TRP=35% in CIL, Table 2) near the first mention.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the replay-savings claims are controlled empirical comparisons, not consequences of the paper's definitions or self-citations.

full rationale

The paper's central claims are empirical measurements: Table 2 and Figure 4 compare final accuracies of configurations at controlled total replay percentages (TRP). RR, CSR, and TRP are defined in Sections 3 and 5.1 as experimental control variables; the result that Seq.+Consol. at TRP=35% matches the no-merge/consol. baseline at TRP=100% is an observed equality, not an algebraic consequence of the TRP formula. Consolidation is defined independently of the accuracy outcome it is claimed to achieve, and no parameter is fitted to the headline replay-reduction number. The only author-overlapping citation is [13] for the replay retrieval strategy; that strategy is applied uniformly across all compared configurations, so it is a controlled component rather than the source of the claimed savings, and it does not make the central result circular. The evaluation-protocol concerns raised in review (no fixed test split, per-configuration tuning on validation, omitted standard deviations, only 3 task-split seeds) are statistical and reporting risks, not circularity, and per the review rules they do not raise the circularity score.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper is empirical; no formal derivation. The central replay-reduction claims rest on a set of tuned hyperparameters (RR, CSR, alpha, LoRA rank) and on the use of the author's own replay-retrieval strategy. No independent evidence is provided for the stability of the pretrained features or the frozen-logit distillation signal.

free parameters (4)
  • Merging coefficient alpha (sequential merging) = tuned per setting (CIL vs CPT, number of tasks)
    Controls the blend of pre-task and post-task weights in sequential merging (theta_t = (1-alpha) theta_{t-1} + alpha theta*_t). The paper tunes it per configuration, so the reported replay savings depend on this choice.
  • LoRA rank r = 8 for tasks of 1-6 classes; 32 for tasks of 10-30 classes
    Selected after testing 4, 8, 16, 32, 64, 128. The choice affects LoRA's regularization and plasticity balance, which is central to the under-regularized regime analysis.
  • Replay ratio (RR) and consolidation step rate (CSR) = e.g., RR=0.25, CSR=0.27 for TRP=45% in the 55% reduction claim
    These knobs directly set the total replay percentage; the headline reductions are achieved by specific tuned values, not by a fixed rule.
  • Replay loss weight / loss coefficients = not stated explicitly
    The paper says it scales the replay loss and uses loss decoupling, but does not report the coefficient values, so replication requires guessing them.
assumptions (3)
  • domain assumption A fixed replay buffer and a uniform retrieval policy provide a fair and effective baseline for all compared methods.
    The paper imports the retrieval strategy from [13] (author's own prior work) and assumes it is equally applicable to all methods compared. Section 3.
  • domain assumption Knowledge distillation from frozen logits of past classes is a stable and sufficient regularizer during task learning and consolidation.
    The replay loss uses DER++-style logit distillation with logit standardization and freezes past-class logits (Section 9). If these logits become stale or miscalibrated, consolidation could reinforce errors.
  • ad hoc to paper The reported numbers come from a held-out test set, and hyperparameter selection uses only validation data.
    No train/test split is described. Section 9 mentions sweeping learning rate and taking the best checkpoint based on validation accuracy, but the final evaluation set is unspecified; the paper's claims implicitly assume a clean evaluation protocol.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scalable Strategies for Continual Learning with Replay." pith.science (2026). https://pith.science/paper/PNIY5VYX

@misc{pith2026250512512,
  author       = {Pith},
  title        = {Pith review of: Scalable Strategies for Continual Learning with Replay},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PNIY5VYX}},
  note         = {Machine review of arXiv:2505.12512}
}
read the original abstract

Future deep learning models will be distinguished by systems that perpetually learn through interaction, imagination, and cooperation, blurring the line between training and inference. This makes continual learning a critical challenge, as methods that efficiently maximize bidirectional transfer across learning trajectories will be essential. Replay is on track to play a foundational role in continual learning, allowing models to directly reconcile new information with past knowledge. In practice, however, replay is quite unscalable, doubling the cost of continual learning when applied naively. Moreover, the continual learning literature has not fully synchronized with the multi-task fine-tuning literature, having not fully integrated highly scalable techniques like model merging and low rank adaptation into a replay-enabled toolset that can produce a unified model in the face of many sequential tasks. In this paper, we begin by applying and analyzing low rank adaptation in a continual learning setting. Next, we introduce consolidation, a phasic approach to replay which leads to up to 55\% less replay samples being needed for a given performance target. Then, we propose sequential merging, an offshoot of task arithmetic which is tailored to the continual learning setting and is shown to work well in combination with replay. Finally, we demonstrate that the developed strategies can operate synergistically, resulting in a highly scalable toolset that outperforms standalone variants.

Figures

Figures reproduced from arXiv: 2505.12512 by the authors.

Figure 1
Figure 1. Visual representation of the impact of changing the re [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. LoRA performance vs. FFT performance across task sizes while keeping total downstream classes fixed at 120 (top). LoRA [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visual representation of continual learning with consolidation. Standard continual learning complements task samples with many [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance of standard replay (no consolidation) versus consolidation (CIL/CPT). Left: mean final accuracy across tasks for each replay-ratio (RR) and consolidation-sample-ratio (CSR) setting, with TRP reporting the total proportion of replayed samples. Note that a CS…
Figure 5
Figure 5. Figure 5: Visual representation of the differences between parallel and sequential merging. Parallel merging trains [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Performances of parallel merging, EMA, and sequential merging while varying the number of downstream tasks, keeping the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Diffract: Spectral View of LLM Domain Adaptation

    cs.LG 2026-08 conditional novelty 6.0 of 10

    During continual pre-training of OLMo 2 models, singular value spectra remain largely fixed while singular vectors change; selectively rewinding low-importance attention heads improves math accuracy by up to 4%.

Reference graph

Works this paper leans on

61 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ss-il: Separated softmax for incremental learning

    Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 824–833, 2021

    work page 2021

  2. [2]

    Distillation scaling laws.arXiv preprint arXiv:2502.08606, 2025

    Dan Busbridge, Amitis Shidani, Floris Weers, Jason Rama- puram, Etai Littwin, and Russ Webb. Distillation scaling laws.arXiv preprint arXiv:2502.08606, 2025

    arXiv 2025

  3. [3]

    Dark experience for general continual learning: a strong, simple baseline.Advances in Neural Information Processing Systems, 33:15920–15930, 2020

    Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline.Advances in Neural Information Processing Systems, 33:15920–15930, 2020

    2020

  4. [4]

    A survey on in-context learning.arXiv preprint arXiv:2301.00234, 2022

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. A survey on in-context learning.arXiv preprint arXiv:2301.00234, 2022

    arXiv 2022

  5. [5]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Ab- hishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, et al. The llama 3 herd...

    arXiv 2024

  6. [6]

    A unified continual learn- ing framework with general parameter-efficient tuning

    Qiankun Gao, Chen Zhao, Yifan Sun, Teng Xi, Gang Zhang, Bernard Ghanem, and Jian Zhang. A unified continual learn- ing framework with general parameter-efficient tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11483–11493, 2023

    work page 2023

  7. [7]

    Tic-clip: Continual training of clip models.arXiv preprint arXiv:2310.16226, 2023

    Saurabh Garg, Mehrdad Farajtabar, Hadi Pouransari, Raviteja Vemulapalli, Sachin Mehta, Oncel Tuzel, Vaishaal Shankar, and Fartash Faghri. Tic-clip: Continual training of clip models.arXiv preprint arXiv:2310.16226, 2023

    arXiv 2023

  8. [8]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen...

    arXiv 2025

Show all 61 references
  1. [9]

    Rusu, and Razvan Pascanu

    Raia Hadsell, Dushyant Rao, Andrei A. Rusu, and Razvan Pascanu. Embracing change: Continual learning in deep neural networks.Trends in Cognitive Sciences, 24(12):1028– 1040, 2020. 11

    2020

  2. [10]

    Parameter-efficient fine-tuning for large models: A comprehensive survey.arXiv preprint arXiv:2403.14608, 2024

    Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey.arXiv preprint arXiv:2403.14608, 2024

    2024 arXiv

  3. [11]

    Hayes, Ronald Kemker, and Christopher Kanan

    Md Yousuf Harun, Jhair Gallardo, Tyler L. Hayes, Ronald Kemker, and Christopher Kanan. Siesta: Efficient online continual learning with sleep.Transactions on Machine Learning Research, 2023. TMLR

    2023

  4. [12]

    Hayes, Giri P

    Tyler L. Hayes, Giri P. Krishnan, Maxim Bazhenov, Hava T. Siegelmann, Terrence J. Sejnowski, and Christopher Kanan. Replay in deep learning: Current approaches and missing biological elements.arXiv preprint arXiv:2104.04132, 2021

    2021 arXiv

  5. [13]

    Watch your step: Optimal retrieval for continual learning at scale.arXiv preprint arXiv:2404.10758, 2024

    Truman Hickok and Dhireesha Kudithipudi. Watch your step: Optimal retrieval for continual learning at scale.arXiv preprint arXiv:2404.10758, 2024

    2024 arXiv

  6. [14]

    Distilling the knowledge in a neural network.ArXiv, abs/1503.02531, 2015

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.ArXiv, abs/1503.02531, 2015

    2015 arXiv

  7. [15]

    Parameter-efficient transfer learning for nlp

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. InProceedings of the 36th International Conference on Machine Learning, pages 2790–279...

    2019

  8. [16]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021

    2021

  9. [17]

    A survey on retrieval- augmented text generation for large language models.arXiv preprint arXiv:2404.10981, 2024

    Yizheng Huang and Jimmy Huang. A survey on retrieval- augmented text generation for large language models.arXiv preprint arXiv:2404.10981, 2024

    2024 arXiv

  10. [18]

    Position: Open-endedness is essential for artificial superhuman intelligence

    Edward Hughes, Michael D Dennis, Jack Parker-Holder, Feryal Behbahani, Aditi Mavalankar, Yuge Shi, Tom Schaul, and Tim Rockt¨aschel. Position: Open-endedness is essential for artificial superhuman intelligence. InProceedings of the 41st International Conference on Machine Lear...

    2024

  11. [19]

    Richter, Quentin Anthony, Timoth ´ee Lesort, Eugene Belilovsky, and Irina Rish

    Adam Ibrahim, Benjamin Th ´erien, Kshitij Gupta, Mats L. Richter, Quentin Anthony, Timoth ´ee Lesort, Eugene Belilovsky, and Irina Rish. Simple and scalable strategies to continually pre-train large language models.arXiv preprint arXiv:2403.08763, 2024

    2024 arXiv

  12. [20]

    Open- clip, 2021

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Han- naneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Open- clip, 2021. If you use this software, please cite it as below

    2021

  13. [21]

    Editing models with task arithmetic.arXiv preprint arXiv:2212.04089, 2022

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic.arXiv preprint arXiv:2212.04089, 2022

    2022 arXiv

  14. [22]

    Unlocking the power of function vectors for characterizing and mitigating catastrophic forgetting in continual instruction tuning

    Gangwei Jiang, Caigao Jiang, Zhaoyi Li, Siqiao Xue, Jun Zhou, Linqi Song, Defu Lian, and Ying Wei. Unlocking the power of function vectors for characterizing and mitigating catastrophic forgetting in continual instruction tuning. InIn- ternational Conference on Learning Repres...

  15. [23]

    Continual pre-training of lan- guage models

    Zixuan Ke, Yijia Shao, Haowei Lin, Tatsuya Konishi, Gyuhak Kim, and Bing Liu. Continual pre-training of lan- guage models. InProceedings of the Eleventh International Conference on Learning Representations (ICLR), 2023

    2023

  16. [24]

    Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A

    James Kirkpatrick, Razvan Pascanu, Neil C. Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska- Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Ku- maran, and Raia Hadsell. Overcoming catastrophic for- g...

    2017

  17. [25]

    McClel- land

    Dharshan Kumaran, Demis Hassabis, and James L. McClel- land. What learning systems do intelligent agents need? complementary learning systems theory updated.Trends in Cognitive Sciences, 20(7):512–534, 2016

    2016

  18. [26]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. InProceed- ings of the 2021 Conference on Empirical Methods in Nat- ural Language Processing, pages 3045–3059, Online and Punta Cana, Dominican Republic, 2021. Association ...

    2021

  19. [27]

    Eurekaverse: Environment curriculum generation via large language mod- els.arXiv preprint arXiv:2411.01775, 2024

    William Liang, Sam Wang, Hung-Ju Wang, Osbert Bastani, Dinesh Jayaraman, and Yecheng Jason Ma. Eurekaverse: Environment curriculum generation via large language mod- els.arXiv preprint arXiv:2411.01775, 2024

    2024 arXiv

  20. [28]

    Loss decoupling for task- agnostic continual learning

    Yan-Shuo Liang and Wu-Jun Li. Loss decoupling for task- agnostic continual learning. InAdvances in Neural Informa- tion Processing Systems, 2023

    2023

  21. [29]

    A survey of in-context reinforcement learning.arXiv preprint arXiv:2502.07978, 2025

    Amir Moeini, Jiuqi Wang, Jacob Beck, Ethan Blaser, Shi- mon Whiteson, Rohan Chandra, and Shangtong Zhang. A survey of in-context reinforcement learning.arXiv preprint arXiv:2502.07978, 2025

    2025 arXiv

  22. [30]

    Task arithmetic in the tangent space: Improved editing of pre-trained models

    Guillermo Ortiz-Jimenez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. InAdvances in Neural Infor- mation Processing Systems, 2023

    2023

  23. [31]

    R+x: Retrieval and execution from everyday human videos.arXiv preprint arXiv:2407.12957, 2024

    Georgios Papagiannis, Norman Di Palo, Pietro Vitiello, and Edward Johns. R+x: Retrieval and execution from everyday human videos.arXiv preprint arXiv:2407.12957, 2024

    2024 arXiv

  24. [32]

    Berg, and Li Fei-Fei

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, San- jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 1...

    2015

  25. [33]

    Rusu, Neil C

    Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Raz- van Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016

    2016 arXiv

  26. [34]

    David Silver and Richard S. Sutton. Welcome to the era of experience. InDesigning an Intelligence. MIT Press, 2025. Preprint available athttps : / / storage . googleapis . com / deepmind - media / Era - of - Experience % 20 / The % 20Era % 20of % 20Experience%20Paper.pdf

    2025

  27. [35]

    A closer look at rehearsal-free continual learning.arXiv preprint arXiv:2203.17269, 2022

    James Seale Smith, Junjiao Tian, Shaunak Halbe, Yen- Chang Hsu, and Zsolt Kira. A closer look at rehearsal-free continual learning.arXiv preprint arXiv:2203.17269, 2022. 12

    2022 arXiv

  28. [36]

    Scaling llm test-time compute optimally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314, 2024

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Ku- mar. Scaling llm test-time compute optimally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314, 2024

    2024 arXiv

  29. [37]

    Improving online continual learning performance and stability with temporal ensembles

    Albin Soutif-Cormerais, Antonio Carta, and Joost Van de Weijer. Improving online continual learning performance and stability with temporal ensembles. InProceedings of The 2nd Conference on Lifelong Learning Agents, pages 828–

  30. [38]

    Logit standardization in knowledge distillation

    Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, and Xi- aochun Cao. Logit standardization in knowledge distillation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15731–15740, 2024

    2024

  31. [39]

    Three scenar- ios for continual learning.arXiv preprint arXiv:1904.07734, 2019

    Gido M van de Ven and Andreas S Tolias. Three scenar- ios for continual learning.arXiv preprint arXiv:1904.07734, 2019

    1904 arXiv

  32. [40]

    Hayes, Eyke H ¨ullermeier, Christopher Kanan, Dhireesha Kudithipudi, Christoph H

    Eli Verwimp, Rahaf Aljundi, Shai Ben-David, Matthias Bethge, Andrea Cossu, Alexander Gepperth, Tyler L. Hayes, Eyke H ¨ullermeier, Christopher Kanan, Dhireesha Kudithipudi, Christoph H. Lampert, Martin Mundt, Razvan Pascanu, Adrian Popescu, Andreas S. Tolias, Joost van de Weij...

    2024 arXiv

  33. [41]

    Lotus: Continual imitation learning for robot manipula- tion through unsupervised skill discovery.arXiv preprint arXiv:2311.02058, 2023

    Weikang Wan, Yifeng Zhu, Rutav Shah, and Yuke Zhu. Lotus: Continual imitation learning for robot manipula- tion through unsupervised skill discovery.arXiv preprint arXiv:2311.02058, 2023

    2023 arXiv

  34. [42]

    V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291, 2023

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandku- mar. V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291, 2023

    2023 arXiv

  35. [43]

    A comprehensive survey of continual learning: Theory, method and application.arXiv preprint arXiv:2302.00487, 2023

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application.arXiv preprint arXiv:2302.00487, 2023

    2023 arXiv

  36. [44]

    Hide-pet: Continual learning via hierarchical de- composition of parameter-efficient tuning.arXiv preprint arXiv:2407.05229, 2024

    Liyuan Wang, Jingyi Xie, Xingxing Zhang, Hang Su, and Jun Zhu. Hide-pet: Continual learning via hierarchical de- composition of parameter-efficient tuning.arXiv preprint arXiv:2407.05229, 2024

    2024 arXiv

  37. [45]

    Scaling pre-training to one hundred billion data for vision language models.arXiv preprint arXiv:2502.07617, 2025

    Xiao Wang, Ibrahim Alabdulmohsin, Daniel Salz, Zhe Li, Keran Rong, and Xiaohua Zhai. Scaling pre-training to one hundred billion data for vision language models.arXiv preprint arXiv:2502.07617, 2025

    2025 arXiv

  38. [46]

    Dualprompt: Com- plementary prompting for rehearsal-free continual learning

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Dualprompt: Com- plementary prompting for rehearsal-free continual learning. InProceedings of the European Conference on Compute...

    2022

  39. [47]

    Learning to prompt for con- tinual learning

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jen- nifer Dy, and Tomas Pfister. Learning to prompt for con- tinual learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages...

    2022

  40. [48]

    Continual learning with low rank adap- tation.arXiv preprint arXiv:2311.17601, 2023

    Martin Wistuba, Prabhu Teja Sivaprasad, Lukas Balles, and Giovanni Zappella. Continual learning with low rank adap- tation.arXiv preprint arXiv:2311.17601, 2023

    2023 arXiv

  41. [49]

    Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V . Le. Self-training with noisy student improves imagenet clas- sification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684– 10695, 2020

    2020

  42. [50]

    Ties-merging: Resolving interference when merging models

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. InAdvances in Neural Information Processing Systems, 2023

    2023

  43. [51]

    What matters for model merging at scale?arXiv preprint arXiv:2410.03617, 2024

    Prateek Yadav, Tu Vu, Jonathan Lai, Alexandra Chronopoulou, Manaal Faruqui, Mohit Bansal, and Tsendsuren Munkhdalai. What matters for model merging at scale?arXiv preprint arXiv:2410.03617, 2024

    2024 arXiv

  44. [52]

    Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities.arXiv preprint arXiv:2408.07666, 2024

    Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xi- aochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities.arXiv preprint arXiv:2408.07666, 2024

    2024 arXiv

  45. [53]

    Continual learners are incremental model generalizers.arXiv preprint arXiv:2306.12026, 2023

    Jaehong Yoon, Sung Ju Hwang, and Yue Cao. Continual learners are incremental model generalizers.arXiv preprint arXiv:2306.12026, 2023

    2023 arXiv

  46. [54]

    Boosting continual learning of vision-language models via mixture-of-experts adapters

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

    2024

  47. [55]

    Language models are super mario: Absorbing abilities from homologous models as a free lunch

    Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. InInternational Con- ference on Machine Learning (ICML), 2024

    2024

  48. [56]

    When scaling meets llm finetuning: The effect of data, model and finetuning method.arXiv preprint arXiv:2402.17193, 2024

    Biao Zhang, Zhongtao Liu, Colin Cherry, and Orhan Firat. When scaling meets llm finetuning: The effect of data, model and finetuning method.arXiv preprint arXiv:2402.17193, 2024

    2024 arXiv

  49. [57]

    C-lora: Continual low-rank adaptation for pre-trained models.arXiv preprint arXiv:2502.17920, 2025

    Xin Zhang, Liang Bai, Xian Yang, and Jiye Liang. C-lora: Continual low-rank adaptation for pre-trained models.arXiv preprint arXiv:2502.17920, 2025

    2025 arXiv

  50. [58]

    Model tailor: Mitigating catastrophic forgetting in multi-modal large lan- guage models.arXiv preprint arXiv:2402.12048, 2024

    Didi Zhu, Zhongyi Sun, Zexi Li, Tao Shen, Ke Yan, Shouhong Ding, Kun Kuang, and Chao Wu. Model tailor: Mitigating catastrophic forgetting in multi-modal large lan- guage models.arXiv preprint arXiv:2402.12048, 2024. 13 Scalable Strategies for Continual Learning with Replay Sup...

    2024 arXiv

  51. [59]

    Training Details For samples from the current task, we use a standard cross- entropy loss. For replay samples, we use a variant of the DER++ [3] objective that replaces raw logit distillation with traditional knowledge distillation (KD) [14] and adds logit standardization [38]...

  52. [60]

    Extra Results CIL CPT N/A Seq. N/A Seq. R 1:20 1:20 PT 1:20 PT 1:20 1.0 73.8 76.0 59.7 58.5 59.9 59.8 0.5 72.8 74.7 59.6 57.7 59.7 59.0 0.25 71.1 72.7 59.1 57.4 59.5 58.6 0.1 63.8 66.1 58.7 56.6 58.9 58.4 Table 3. Reducing the replay ratio for sequential merging and a no- merg...

  53. [61]

    Relation to Reinforcement Learning Reinforcement learning (RL) is, fundamentally, continual learning. Off-policy RL agents store past interactions in a replay buffer and sample mini-batches of these experiences to break the temporal correlations in continuously gener- ated dat...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.