acceptodds
Under review as a conference paper at ICLR 2027

Replay-Gap Thresholds: Sharp Burst Guarantees and Coupled Retention

Abstract

Sparse replay can retain training examples even when their update frequency vanishes, but the placement of updates matters. We prove a sharp worst-case replay-gap threshold for constant-step online logistic regression on any separable interfering pair. For a fixed burst of rare updates, let be its total scalar logistic response from margin , let in coordinates , and set . Under an explicit sufficient step-size bound, every replay-time ratio sequence confined eventually to a fixed interval guarantees eventual all-time retention from every finite initialization. Geometric schedules above the threshold establish worst-case sharpness; equality remains open. The proof uses a common invariant region controlled by the derivative of the full burst response. We also extend the centered-map method to any finite number of orthogonal common examples coupled through one interfering rare example, obtaining an exact geometric sign criterion in the stated exposure regime and a computable local guarantee for arbitrary persistent ratio variation. In this family, two schedules whose rare-update counts differ by at most one at every prefix can have opposite retention outcomes, robustly under small bounded shared log-time jitter. The results concern logarithmically sparse replay and raw training margins, not population accuracy.

Primary area:
learning theory
Keywords:
continual learning, replay scheduling, online logistic regression, retention, dynamical systems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HHf7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1(−1)
Soundness
2
Read review
Paper summary

This paper investigates how replay timing affects sample retention in constant-step online logistic regression. The authors establish a sharp worst-case replay-gap threshold for separable interfering pairs under fixed-size bursts and prove eventual all-time retention under subcritical replay-time ratios. The analysis is extended to a coupled multi-example family, providing a geometric sign criterion and a local guarantee under persistent ratio variation. Furthermore, the paper demonstrates that replay schedules with nearly identical cumulative update counts can lead to opposite retention outcomes. The results are supported by theoretical proofs and a limited numerical evaluation of the derived sufficient conditions.

Critical strengths

The paper provides a well-structured theoretical analysis of replay-gap thresholds in online logistic regression. The authors present a set of formally stated results with extensive proofs that examine the effects of replay timing and burst scheduling on sample retention. The paper also includes a numerical evaluation of the derived sufficient conditions in Table 1. Overall, the theoretical development appears technically rigorous and constitutes a strength of the submission.

Critical weaknesses
  1. Limited Significance and Practical Relevance. The paper does not sufficiently establish the broader relevance of its theoretical results to continual learning. The analysis focuses on logarithmically sparse replay schedules, but it remains unclear which practical continual learning scenarios motivate such schedules or how the derived replay-gap thresholds could guide replay strategies. Furthermore, Section 2 defines retention solely in terms of positive training margins, while Section 8 acknowledges that the normalized rare margin can still converge to zero. The paper does not establish how this notion of retention relates to forgetting, generalization, or downstream task performance. Consequently, their significance for understanding practical continual learning remains insufficiently justified.

  2. Limited Generalizability and overly idealized assumptions. Although the paper extends its analysis to varying replay-time ratios and a coupled multi-example family, the theoretical framework still relies on restrictive assumptions, including fixed-feature logistic regression, deterministic replay schedules, and specific sample geometries. In particular, the multi-example extension in Theorem 3 and Appendix F requires orthogonal common examples and a single interfering rare example. As acknowledged in Section 8, the results do not cover arbitrary larger datasets, multiple rare directions, stochastic replay laws, or representation learning. While these simplifications facilitate rigorous theoretical analysis, the paper provides insufficient justification that the identified replay-timing phenomena capture more general mechanisms of forgetting or replay effectiveness in continual learning systems.

Questions for authors
  1. Could the authors clarify which practical continual learning scenarios motivate the logarithmically sparse replay schedules considered in this paper? How could the derived thresholds inform the design or analysis of practical replay strategies?

  2. To what extent do the theoretical conclusions depend on the restrictive sample geometries and deterministic replay assumptions? Is there evidence or a theoretical argument suggesting that the identified phenomena persist in more general continual learning settings?

  3. The paper primarily relies on mathematical derivations, with limited intuitive explanations or visual illustrations of its main findings. Could the authors provide more intuitive explanations or visualizations to help readers understand the theoretical results? For example, a comparison of replay schedules and their corresponding margin trajectories in Section 6 would help illustrate why schedules with nearly identical cumulative update counts can lead to opposite retention outcomes.

  4. The Reproducibility Statement explicitly refers to an accompanying Python script for reproducing the numerical results in Table 1. However, there are no submission materials or anonymous link for code. Could the authors clarify its availability and provide the referenced artifact?

Reviewer_nTFYOpenReview ↗
Rating
1
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates the issue of replay intervals in continual learning. The authors establish an analyzable logistic regression framework and analyze the sharp burst guarantees within this framework.

Critical strengths

This paper establishes an exact theory for the two-sample case and generalizes it to the multi-sample case.

Critical weaknesses

The paper isn't presented very well.

The theory developed in this paper is too narrow and constrained by the model; it fails to demonstrate transferability from both theoretical and experimental perspectives. The research findings are not strong enough to be accepted by a top-tier conference.

Questions for authors

Based on the analysis in this article, are there any insights that can be applied to practical replay methods?

Reviewer_D34POpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how the timing of sparse replay affects retention in online logistic regression. It establishes a sharp worst-case replay-ratio threshold for a separable interfering pair under stated step-size conditions, extends parts of the analysis to a structured multi-example setting, and constructs replay schedules whose cumulative counts differ by at most one but have opposite retention outcomes.

Critical strengths

The paper makes a precise distinction between replay frequency and replay timing. In particular, the bounded-count-discrepancy construction provides an interesting demonstration that nearly identical replay budgets need not imply equivalent retention. The treatment of persistently irregular timing also goes beyond analyzing a single geometric schedule. The paper clearly states the scope of its guarantees and distinguishes raw training-margin retention from generalization.

Critical weaknesses

My main concern is that the paper provides no training experiments to assess whether the identified timing effects persist in realistic learning settings. Table 1 numerically evaluates the sufficient-condition formulas, but it does not test the learning dynamics on data. The analysis relies on highly structured assumptions, including fixed features, a separable interfering pair or orthogonal common examples, and a single rare example. Consequently, it remains unclear how informative the results are for replay scheduling on real datasets. I appreciate that this is a theoretical contribution, and I am not asking the authors to establish their exact threshold for general neural networks. However, the central qualitative finding—that replay placement can substantially change retention even under closely matched replay counts—is directly testable. Empirical evidence for this finding would substantially strengthen the paper’s significance. I would particularly encourage the following experiments:

  1. Compare replay schedules on a real dataset under matched budgets. Use a continual or imbalanced classification setting, with frequent updates from common classes or tasks and sparse replay from a rare class or earlier task. Compare temporally distributed replay against concentrated bursts, keeping the replay buffer, replay examples, total replay updates, optimizer, and total training steps matched. Where feasible, construct schedules with bounded prefix-count discrepancy, mirroring the theoretical comparison.
  2. Evaluate retention throughout training. Since the paper’s guarantee concerns eventual all-time retention, reporting only final accuracy could miss the relevant phenomenon. Please report rare-task or rare-class performance over time, including measurements immediately before and after replay bursts, together with common-task performance. Training margins would connect the experiments to the theory, while held-out accuracy would establish whether the effect extends beyond the replayed examples.
  3. Test whether the effect survives less restrictive models and data geometry. A useful progression would be logistic regression on fixed features extracted from real data, followed by a small neural network with learned representations. Varying burst size and replay spacing at matched budgets would help determine whether the predicted disadvantage of concentrated replay persists outside the analytically tractable setting. Multiple seeds and uncertainty estimates should be included.

These experiments need not involve large language models or large-scale training. A carefully controlled study on a standard real-world dataset would already help distinguish a broadly relevant scheduling effect from a phenomenon that depends strongly on the paper’s particular assumptions. If the qualitative effect weakens or disappears outside those assumptions, that would also be valuable evidence for clarifying the scope of the contribution.

Questions for authors
  1. Can the opposite retention outcomes under closely matched replay counts be reproduced on real data?
  2. Does the effect remain visible in held-out performance, rather than only in the margins of individual replayed training examples?
  3. Which aspects of the theory appear most important empirically: interference geometry, fixed representations, constant step size, or the sparsity of replay?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.