acceptodds
Under review as a conference paper at ICLR 2027

AllocAdapt: Adaptive Training Budget Allocation for Continual Learning

Abstract

Recent class-incremental learning (CIL) methods achieve strong performance by adapting pre-trained models with lightweight adapters and low-rank updates. However, their parameter efficiency does not necessarily translate into efficient training, as fixed schedules overlook differences in optimization demand across tasks and stages. We propose AllocAdapt, an adaptive update allocation framework that jointly addresses two complementary decisions: how much to update and where to allocate updates. HowMuch combines probe statistics with historical calibration to determine task- and stage-level training horizons under budget constraints. Within these horizons, Where allocates update opportunities through parameter-update routing or class sampling, according to each learner's optimization structure. Experiments on three recent parameter-efficient CIL methods and three image classification benchmarks show that AllocAdapt reduces training iterations by up to 80% while largely preserving average incremental accuracy relative to the original training procedures.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual Learning, Class-incremental learning, Parameter-efficient adaptation, Adaptive update allocation, Training efficiency, Pre-trained models.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_AcWgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors observe that existing parameter-efficient continual learning methods focus on reducing the number of trainable parameters while overlooking the impact of training iterations and optimization strategies on training efficiency. To address this limitation, they propose an adaptive training budget allocation framework for parameter-efficient continual learning. The proposed method significantly reduces the number of training iterations with only a minor loss in accuracy.

Critical strengths
  1. The authors investigate parameter-efficient continual learning from the perspectives of reducing training iterations and update allocation. This novel perspective offers significant research value and has far-reaching implications for future research.
  2. The experimental results are impressive, demonstrating that the proposed method reduces the number of training iterations required by existing parameter-efficient continual learning methods.
Critical weaknesses
  1. I believe the proposed ‘HowMuch’ component has strong practical value. However, for the ‘Where’ component, the authors design three different allocation strategies for three distinct CL methods, indicating that ‘Where’ relies heavily on the specific architecture and optimization structure of each method. This raises concerns about whether ‘Where’ can truly serve as a generalizable training budget allocation framework.
  2. A reduction in the number of training iterations does not necessarily translate into a reduction in overall training time. The proposed framework introduces additional computational overhead from probe statistics, loss-gradient routing, class sampling, or classifier-head selection, which may increase the time required per epoch.
Questions for authors
  1. As mentioned in Weakness 1, the ‘Where’ component relies on the specific architecture and optimization structure of each continual learning method to allocate update opportunities. For example, EWC-LoRA (ICLR 2026), a regularization-based continual learning method, does not have the dual-branch structure of CL-LoRA or the classifier-alignment interfaces used in MOS/TUNA. Would AllocAdapt remain effective when applied to such methods, and how much method-specific modification would be required? Does this imply that a new Where strategy must be designed whenever AllocAdapt is extended to a new continual learning method?
  2. Regarding Weakness 2, could the authors report the overall training time of AllocAdapt and the original methods under the same hardware conditions, including the additional computational overhead introduced by g probe statistics, loss-gradient routing, class sampling, or classifier-head selection? This would help determine whether the substantial reduction in training iterations translates into meaningful training speedups in practice.
  3. Equation (2) assumes that a weighted combination of R, G, and I can represent training demands. How can the authors demonstrate that the demand score defined in Equation (2) accurately reflects the actual optimization needs?
4 comments
AuthorsOpenReview ↗

A Gentle Request for Feedback on Our Rebuttal

Dear Reviewer,

Thank you again for your constructive feedback. We have provided point-by-point responses to all the questions and concerns raised in your review and conducted additional experiments. We have also revised the manuscript to incorporate these results and clarifications and uploaded the updated PDF to OpenReview.

We would appreciate it if you could review our responses and the revised manuscript and let us know whether any concerns remain. We would be happy to provide further clarification and would welcome your updated assessment in light of these additions.

Thank you for your time and consideration.

Best regards, The Authors

AuthorsOpenReview ↗

Shared structure of Where and evidence for adaptive allocation

Thank you for your questions about the generality of Where. We would like to make explicit which elements are shared and how they connect to the different training interfaces.

1. Shared allocation structure

Where follows a common decision structure: identify allocation units in the native training process, obtain optimization feedback, and assign training opportunities under the horizon selected by HowMuch and the constraints of that interface.

LearnerAllocation decisionTraining feedbackExecution constraint
CL-LoRAWhen to switch from joint training to separated CE/KD routesCE/KD improvement and KD reboundSwitching occurs within the assigned horizon and admissible decision interval
MOSHow to distribute synthetic features across classesClass-gradient statisticsFixed total feature budget and balanced target objective
TUNAHow frequently to update classifier parameter blocksBlock-gradient statisticsA fixed number of active blocks per routed step, with coverage scheduling

The common contribution is this organization of training allocation: feedback informs the distribution of work within an assigned horizon. The learner-specific interface determines how that decision is executed. All three integrations retain the base architecture and inference procedure.

For another learner, the integration requires specifying its allocation units, observable training signals, and admissible actions. For a regularization-based learner, existing parameter groups and gradients of the native regularized objective offer a possible interface. Its effectiveness would need to be evaluated for that integration.

2. Interpretation and empirical role of the demand score

Regarding the demand score, RR captures residual loss relative to the probe's starting loss, GG captures continued improvement, and II captures recent loss increases. These signals distinguish training trajectories with similar endpoint losses but different ongoing progress or instability. Historical normalization interprets them relative to earlier tasks of the same stage.

The component comparisons provide evidence about the resulting allocation decisions. Full has the highest reported Avg and Final in all four groups in Table 2 at iteration budgets comparable to the component variants. For MOS specifically, adding HowMuch to Where-only improves Final on both evaluated datasets while using fewer iterations. The recorded trajectories additionally connect calibrated demand to executed horizons.

3. Practical efficiency including allocation work

Practical efficiency gains are observed across all nine learner–dataset settings, with measured training-time reductions of 13.97–54.80%, including probing and allocation work.

The supplementary revision summary identifies the corresponding additions. We hope these clarifications make the shared allocation structure and its empirical support easier to assess.

Thank you for the author's reply. The concerns about generalizability are not fully resolved. For regularization-based learners, the author only proposed one possible integration method without actual verification. For Eq.(2), I understand the design motivation, but the rationality and reliability of the score still lack direct verification. Therefore, I'm keeping my socre.

Thank you for your follow-up. We would like to clarify the motivation for selecting the three base learners and the experimental evidence supporting demand-guided allocation.

1. Competitive base learners and the scope of generalization

AllocAdapt is an allocation framework built on existing continual learners. Its contribution is to determine how much training to perform and where to allocate updates, while preserving the recognition capabilities of the underlying learners. Implementing the framework on established methods therefore directly serves our research objective.

We selected CL-LoRA, MOS, and TUNA because they combine competitive recognition performance with different optimization interfaces. Their published benchmark results illustrate the relevance of this choice. For example, the following results are reported in Table 1 of the TUNA paper. Each entry denotes Avg / Final accuracy (%).

MethodCIFAR-100ImageNet-RImageNet-A
EASE91.51 / 85.8081.74 / 76.1765.34 / 55.04
SLCA92.49 / 88.5581.17 / 77.0068.66 / 58.74
SSIAT93.52 / 90.0783.20 / 78.8570.83 / 62.23
RanPAC94.00 / 90.6282.98 / 77.9469.32 / 61.82
MOS93.30 / 89.2582.96 / 77.9367.08 / 56.22
TUNA94.44 / 90.7484.22 / 79.4273.78 / 64.78

Source: TUNA, Table 1. These experiments use 20 tasks for CIFAR-100 and 10 tasks for ImageNet-R and ImageNet-A.

The comparison with the regularization-based learner mentioned in your review is also informative. Table 2 of the EWC-LoRA paper reports the following results under its own 10-task evaluation protocols:

MethodCIFAR-100ImageNet-RImageNet-A
CL-LoRA92.25 / 87.6585.20 / 78.7270.76 / 57.62
EWC-LoRA92.27 / 87.9178.95 / 72.8668.33 / 59.89

Source: EWC-LoRA, Table 2. Comparisons should be made within each table because the source papers use different experimental protocols.

In particular, CL-LoRA exceeds EWC-LoRA by 6.25 percentage points in Avg and 5.86 points in Final on ImageNet-R, and by 2.43 points in Avg on ImageNet-A. These results demonstrate the competitiveness of the selected learners relative to a recent regularization-based approach on these benchmarks, motivating the study of their training efficiency.

Beyond baseline accuracy, our three implemented integrations cover loss-gradient routing in CL-LoRA, class-sample allocation in MOS, and classifier-block updates in TUNA. All three are evaluated on three datasets. The shared contribution is the probe-and-calibrate allocation framework, instantiated through each learner’s available update interface. This provides experimental evidence across several distinct interfaces within the evaluated setting.

2. Evidence for demand-guided horizon selection

Equation (2) combines complementary training signals: residual loss, late-probe progress, and instability. These signals are calibrated against earlier tasks within the same training stage to guide horizon selection.

The component experiments provide evidence across all four learner–dataset comparisons in Table 2 of our manuscript. Full achieves higher Avg and Final accuracy than Where-only while using the same or fewer iterations. Both variants enable Where, so this comparison evaluates the additional contribution of adaptive horizon selection.

Learner / DatasetWhere-only Avg / FinalFull Avg / FinalFinal improvement
CL-LoRA / CIFAR-10089.679 / 83.4790.160 / 84.56+1.09
CL-LoRA / ImageNet-R81.556 / 73.4581.821 / 74.63+1.18
MOS / CIFAR-10092.826 / 88.6692.852 / 88.88+0.22
MOS / ImageNet-R84.110 / 78.6084.129 / 78.63+0.03

Accuracy is reported in percent; improvements are in percentage points.

For example, MOS uses the score in Eq. (2). Adding HowMuch improves both accuracy metrics on CIFAR-100 and ImageNet-R while reducing iterations. The corresponding CL-LoRA comparisons also show improvements, using its learner-specific demand mapping.

Together, these results provide component-level evidence for the practical utility of demand-guided horizon selection. MOS offers an example directly relevant to the controller using Eq. (2), while CL-LoRA provides complementary evidence for the broader probe-and-calibrate allocation principle.

We have revised the manuscript to include training-time measurements, multi-seed statistics, and additional implementation details. A summary of the manuscript revisions is available in the supplementary material.

Reviewer_3sseOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AllocAdapt adapts the training effort of existing parameter-efficient class-incremental learners. HowMuch uses probe losses and historical calibration to set task/stage horizons; Where reallocates gradient paths, synthetic class samples, or classifier-head updates. Experiments on CL-LoRA, MOS, and TUNA across three benchmarks report 15.10–80.04% fewer training iterations, generally with small average-accuracy changes, plus component comparisons against shortened fixed schedules.

Critical strengths
  1. The framework addresses a practical cost overlooked by parameter counts. Applying training allocation to three different learners demonstrates useful breadth without changing their inference procedures. Seven of nine comparisons remove over half the iterations while keeping average incremental accuracy within 0.5 points of the original learner (Table 1).
  2. The controls address important confounds. Probe updates count toward the budget, and Table 2 compares components at similar iteration counts. MOS also corrects its sampling objective for unequal allocation. These choices make the evidence more informative than a comparison against full-length schedules alone.
Critical weaknesses
  1. Iteration savings do not establish the practical compute benefit. The metric adds adapter-training and synthetic-feature classifier-alignment iterations without cost weighting (Section 4.1; Appendix B.1), although their costs can differ substantially. No wall-clock or FLOP measurements include controller overhead. The stated iteration reductions are meaningful, but they do not show how much actual training computation the framework saves.
  2. The additional value of adaptation is weakly resolved. Against matched shortened schedules, average-accuracy gains are only 0.11–0.48 points (Table 2); the paper gives neither the number of runs nor uncertainty estimates. TUNA has no component ablation, and complete numerical controller settings and their selection procedure are not supplied. This leaves unclear whether the added control machinery reliably improves on a well-chosen fixed budget.
Questions for authors
  1. Cost clarification: What are the stage-wise elapsed times or FLOPs for Original, budget-matched Static, and Full, including probing and allocation overhead on identical hardware?
  2. Evidence needed for assessment: Can you provide run counts, paired accuracy differences with uncertainty, and a TUNA component comparison at matched budgets? How were the reference horizons, probe lengths, bounds, and controller coefficients selected without test-set feedback?
2 comments
AuthorsOpenReview ↗

Training-time evidence across all evaluated settings

Thank you for emphasizing that iteration counts alone do not capture the unequal costs of training operations. We would like to summarize the measured training-time reductions across all evaluated settings.

1. Training-time reductions across all nine settings

The following reductions are reported in the revised manuscript under the timing protocol in Appendix B.3:

LearnerCIFAR-100ImageNet-RImageNet-A
CL-LoRA45.50%15.15%13.97%
MOS44.28%45.67%54.80%
TUNA45.70%43.36%40.46%

Training time decreases in all nine learner–dataset combinations. Each Original/Ours pair uses the same physical GPU. The measured interval includes probing, controller and allocation work, and the applicable feature-statistics, classifier-alignment, and prototype operations. These results therefore account for the additional allocation work when measuring practical training savings.

2. Stage-level comparison including Static

For the stage-level comparison requested in your review, MOS/ImageNet-R provides the following example:

Training time, minutesOriginalStaticFull
Adapter training31.4815.7215.34
Classifier alignment13.146.557.01
Complete measured training interval48.8426.4526.54

The complete interval also includes the applicable feature-statistics and prototype operations. Full retains a 45.67% reduction relative to Original, with a measured duration approximately 0.09 minutes above Static.

3. Accuracy and run statistics

The allocation-quality evidence covers all four component comparisons in Table 2: Full improves both Avg and Final over Static with the same or fewer iterations. Separately, Table 1 and Appendix B.6 report the means, sample standard deviations, and individual results for ten Original and ten Ours runs per setting.

The supplementary material includes a revision summary. We hope the overall timing results and the stage-level breakdown make the practical cost evidence clearer.

Thanks for the responses. My concerns have been resolved. Therefore, I maintain my positive score.

Reviewer_H96HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1(−1)
Soundness
2
Read review
Paper summary

This paper proposes AllocAdapt, a framework for adaptive training budget allocation in continual learning. It has two components: ‘HowMuch’ determines the training horizon, while ‘Where’ decides how to allocate updates within that budget. Experiments show that AllocAdapt substantially reduces training iterations while largely maintaining accuracy.

Critical strengths
  1. Training efficiency is a practical motivation for continual learning. The proposed approach complements existing methods by improving how their training budgets are used.
  2. The paper is well organized. The separation between ‘HowMuch’ and ‘Where’ makes the main idea clear.
  3. Experiments on three base methods and three datasets show substantial reductions in training iterations with limited loss in accuracy.
Critical weaknesses
  1. It is unclear how the reference budgets and probe settings are chosen. For example, how long does the probe train, and how much data does it use? Which settings are shared across methods and datasets, and which are tuned separately? How to choose an appropriate budget?
  2. The three base methods use different allocation strategies. It is unclear how much can be reused for a new method and how much requires additional design and tuning. More details would help assess how easy the framework is to apply elsewhere. If each new method needs a separately designed allocation strategy, this could limit the framework’s practical use.
  3. It would be useful to test tasks with different levels of difficulty, such as fine-grained classification or different numbers of classes per task. Can the probe still estimate training needs reliably without retuning?
  4. In Table 2, Static less iteration budgets already achieve nearly the same iteration savings, while the full method gives relatively small accuracy gains in several cases. Therefore, it seems the main contribution of this paper comes from fewer training iterations rather than the method design. Comparisons with other simple budget-allocation or stopping strategies under the same iteration budget would help show the value of the proposed method.
  5. Since the improvement from Static to Full is marginal in some settings (<0.5%), results across several budgets and standard deviations across runs would also be helpful.
Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

A Gentle Request for Feedback on Our Rebuttal

Dear Reviewer,

Thank you again for your constructive feedback. We have provided point-by-point responses to all the questions and concerns raised in your review and conducted additional experiments. We have also revised the manuscript to incorporate these results and clarifications and uploaded the updated PDF to OpenReview.

We would appreciate it if you could review our responses and the revised manuscript and let us know whether any concerns remain. We would be happy to provide further clarification and would welcome your updated assessment in light of these additions.

Thank you for your time and consideration.

Best regards, The Authors

AuthorsOpenReview ↗

Further details on configuration choices and the complete Static comparisons

Thank you for your questions about the settings and the benefit of adaptation. We would like to clarify how the configured budgets relate to the online decisions and summarize the complete fixed-budget evidence.

1. Shared procedure and configuration choices

The shared HowMuch procedure collects training signals during a probe, calibrates them against earlier tasks of the same stage, and maps the calibrated demand to an executable horizon. Probe updates count toward that horizon. Tables 4–5 specify the numerical settings: MOS and TUNA use three-epoch classifier-alignment probes, their Adapter probe lengths depend on the configuration, and CL-LoRA uses 20% of the original incremental-task horizon.

The budget range specifies the available choices; the current probe and historical calibration determine the task-specific allocation. The recorded ImageNet-R horizons illustrate this behavior: CL-LoRA executes 70.5–89.7% of the original task iterations, MOS executes 14–16 classifier-alignment epochs, and TUNA executes 7–9. Figure 5 connects these executed horizons to the calibrated demand.

2. Complete comparison with Static

For the comparison with uniform shortening, the complete results from Tables 2 and 6 are summarized below. Accuracy gains are Full minus Static, in percentage points.

LearnerDatasetStatic iterationsFull iterationsAvg gainFinal gain
CL-LoRACIFAR-1006,7946,794+0.477+1.14
CL-LoRAImageNet-R12,41212,281+0.262+1.25
MOSCIFAR-10018,99018,937+0.110+0.23
MOSImageNet-R23,74523,549+0.353+0.09

Both accuracy metrics improve in all four comparisons without an increased iteration budget. HowMuch-only and Where-only further separate adaptive training length from adaptive update allocation, and Full has the highest reported Avg and Final in every group.

3. Run variability

For variability in the main comparison, Table 1 and Appendix B.6 report ten Original and ten Ours runs per learner–dataset setting, including means, sample standard deviations, and individual results.

The supplementary revision summary identifies these additions. We hope this presentation makes the configuration choices and the evidence across settings easier to assess.

Reviewer_Z9bMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper tackles class incremental learning from the efficiency perspective and proposes to use HowMuch and Where components to determine how much to update in the network and which parts. Experiments on multiple benchmarks show improvement in training efficiency over baselines.

Critical strengths
  1. The paper motivates the problem well, and efficient updates of network parameters are a timely and significant problem to address.
  2. Experiments on multiple benchmarks show substantial schedule reductions with relatively small accuracy changes.
Critical weaknesses
  1. The major concern about this paper is it's contribution to the continual learning literature. It is unclear what the main novel contributions of the paper are, given that its main components, HowMuch and Where extensively utilize prior methods such as MOS.
  2. The method is designed for CIL where tasks are disjoint, but that might not be the case in most real-world applications. Additionally, tasks might still have overlapping data, even in CIL. How do HowMuch and Where determine how much and where to allocate based on this? From my current understanding, the budget is fixed.
  3. Another issue is the choice of the efficiency metric used in the paper. Simply reducing the number of iterations does not necessarily show improved training efficiency; the paper needs to include wall-clock time FLPOS to understand how much compute is being spent per iteration.
  4. The main table compares AllocAdapt against each learner’s full original schedule. This shows that the reference procedures are often overtrained, but it does not establish that the proposed adaptive controller is necessary to obtain the savings.
  5. All experiments use image-classification CIL, frozen ViT-B/16 backbones, disjoint classes, relatively regular task structures, and three recent modular methods. Therefore, broader claims about modular or parameter-efficient CIL should be limited to the evaluated visual CIL setting.
  6. The writing and Figure 1 of the paper can be improved. Currently, Figure 1 does not seem to add much to the understanding of this paper, and many graphs in the figure are not grounded in any experiments in the paper.
Questions for authors

Please see the weaknesses for my questions.

2 comments
AuthorsOpenReview ↗

A Gentle Request for Feedback on Our Rebuttal

Dear Reviewer,

Thank you again for your constructive feedback. We have provided point-by-point responses to all the questions and concerns raised in your review and conducted additional experiments. We have also revised the manuscript to incorporate these results and clarifications and uploaded the updated PDF to OpenReview.

We would appreciate it if you could review our responses and the revised manuscript and let us know whether any concerns remain. We would be happy to provide further clarification and would welcome your updated assessment in light of these additions.

Thank you for your time and consideration.

Best regards, The Authors

AuthorsOpenReview ↗

Further clarification of the contribution and evidence across settings

Thank you again for reviewing our work. We would like to clarify the relationship between AllocAdapt and the underlying learners, and explain how the experimental comparisons assess its contribution.

1. Contribution of the allocation framework

AllocAdapt addresses the allocation of training effort within existing CIL learners. The underlying methods provide the representations, losses, and native training operations; AllocAdapt introduces feedback-driven decisions about how long these operations should run and how training opportunities should be distributed. HowMuch uses current probes and earlier-task statistics to select task-dependent horizons, while Where allocates updates within those horizons. Evaluating the framework on existing learners therefore allows us to study these additional decisions while retaining their architectures and inference procedures.

For example, MOS provides adapter training and classifier alignment. AllocAdapt adds adaptive stage horizons and gradient-guided class allocation within a fixed feature budget. The balanced target objective is retained through the corresponding loss correction. This illustrates the distinction between the base learner and the proposed allocation policy.

2. Evidence across all evaluated settings

The empirical evidence covers all nine combinations of CL-LoRA, MOS, and TUNA with CIFAR-100, ImageNet-R, and ImageNet-A. Across these settings, AllocAdapt reduces iterations by 15.10–80.04% while largely preserving average incremental accuracy. The reported timing experiments additionally show training-time reductions of 13.97–54.80%, including probing and allocation work.

3. Evidence beyond uniformly shortening training

The fixed-budget evidence extends across all four comparisons in Table 2. In every group, Full achieves the highest reported Avg and Final while using the same or fewer iterations than Static. The gains over Static range from 0.110 to 0.477 percentage points in Avg and from 0.09 to 1.25 points in Final. These comparisons directly examine whether adaptive allocation improves the use of a reduced training budget.

The revised manuscript and supplementary revision summary provide the corresponding settings and results. We hope this clarifies both the framework contribution and the evidence supporting it.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.