acceptodds
Under review as a conference paper at ICLR 2027

Reachable Adaptation under Delayed Feedback: Budgeted Partial Updates in Nonstationary Streams

Abstract

A deployed predictor may receive labels only after delay while each maintenance cycle can update only a few model blocks. We ask which part of the current error is reachable from the model already in service. We formulate current risk over an anchored reachable family, separating mismatch from delayed or replayed evidence, localization from frozen directions, adaptive route selection, and restricted estimation. Under stated conditional-law and second-moment conditions, projection geometry characterizes localization; a coherence analysis compares an observable risk-change route with a joint-gain population oracle; and replay-aware bounds expose the roles of staleness, effective sample size, and selected dimension. Drift-Decomposed Adaptation (DDA) instantiates this view by differencing anchored block gradients across two delayed windows and refitting only selected blocks. Controlled studies probe these mechanisms and their boundary regimes. On measured Appliances data, DDA has RMSE 0.535 versus 0.537 for Full with 66.7% fewer block updates. On a measured Fjord5G segment, it reduces RMSE by 10.1% relative to GradMag at the same one-block budget. These offline chronological evaluations illustrate regime-dependent resource–error tradeoffs rather than universal selector dominance.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
nonstationarity, delayed feedback, partial model updates, continual learning, test-time adaptation, modular learning, resource-constrained adaptation, drift

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_iohKOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates model adaptation in nonstationary data streams under delayed supervision and limited parameter update budgets. To address this problem, the paper proposes Drift-Decomposed Adaptation (DDA), which selects parameter blocks using temporal differneces between gradients computed over historical and recent windows. This paper conducts experiments on synthetic data, Appliances and Fjord5G to evaluate DDA.

Critical strengths
  1. This paper considers delayed supervision and constrained update budgets, which are important challenges in practical nonstationary learning systems.
  2. The anchored reachable-risk framework provides a useful way to distinguish distributional changes from prediction errors that can be corrected.
Critical weaknesses
  1. The theoretical results rely on restrictive assumptions, and the resulting risk bounds may be vacuous in practically relevant nonstationary settings. This limits the informativeness of Theorem 4 and Corollary 3.
  2. The experimental results do not demonstrate the effectiveness or robustness of DDA. On Appliances, DDA achieves an RMSE of 0.5354, which is slightly worse than GradMag and Recent gain. Furthermore, DDA does not achieve the best performance on any of the four additional Fjord5G segments reported in Appendix Table 18.
Questions for authors

See weaknesses.

Reviewer_eAY2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies model adaptation in nonstationary streams under delayed supervision and a hard budget on the blocks updated at each maintenance step. It defines an anchored reachable hypothesis class and separates current-risk degradation into distribution mismatch, localization due to frozen blocks, and selection/estimation effects. The proposed Drift-Decomposed Adaptation (DDA) selects blocks using differences between gradients evaluated on reference and recent delayed windows at the same deployed anchor, followed by restricted refitting.

Critical strengths
  • The problem is practically relevant. Separating limitations imposed by stale information from those imposed by restricted parameter movement provides a useful perspective on budgeted model maintenance.

  • The analysis accounts for correlations between update directions. The projection characterization explains why omitted changes can sometimes be compensated by active blocks, while the joint-gain oracle avoids assuming that individual block gains are additive.

  • The authors include persistent-error failure cases, heterogeneous cross-segment results, and timing measurements that distinguish reduced update volume from actual runtime savings.

Critical weaknesses
  • Equation (7) takes a supremum of the hybrid/current risk difference over an unrestricted reachable class. Consider one-dimensional regression with X∼N(0,1)X \sim\mathcal N(0,1), equal noise variance, and old/current regression coefficients a≠ba\neq b. For hw(X)=wXh_w(X)=wX, Ra(w)−Rb(w)=2(b−a)w+a2−b2R_a(w)-R_b(w)=2(b-a)w+a^2-b^2, so the discrepancy is infinite when ww is unrestricted.

  • DDA significantly improves over GradMag on the headline Fjord5G segment at (K=1), but the broader results are mixed. On Appliances, its linear RMSE is slightly worse than GradMag and Recent gain. It does not attain the lowest reported RMSE on any of the four additional Fjord5G segments. In the neural experiments, DDA is worse than GradMag on Appliances and nearly identical on Fjord5G. These results support the potential value of partial updating more clearly than they establish an advantage of the proposed routing score.

  • With exactly K out of M blocks updated, the reduction in block-update count follows directly from the imposed budget. All blocks must still be scored, and the reported (M=128) timing is nearly unchanged: 17.825 ms for DDA versus 17.887 ms for Full.

Questions for authors
  • Can the authors provide a non-vacuous current-risk guarantee under nonzero concept drift, using a precisely specified bounded or localized reachable class compatible with the implemented refitting procedure?

  • Why were the headline Fjord5G segment and budget selected? Would an aggregate evaluation across predefined segments and chronologically selected hyperparameters support the same conclusion?

  • Which deployment resource is represented by block-update count, and can its savings be measured directly?

Reviewer_2s5aOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies budgeted partial updates under delayed feedback and replay. An anchored reachable class preserves omitted blocks at their deployed values. DDA selects blocks using gradient differences across two delayed windows and performs a restricted refit. The analysis covers hybrid-to-current mismatch, localization bias, routing relative to a joint-gain oracle, and restricted estimation. Controlled mechanism studies and retrospective measured-stream evaluations demonstrated the predicted gains and boundary regimes.

Critical strengths
  1. Interesting problem setting: I agree that detecting distributional change does not determine which blocks should receive an update. For practical model maintenance, it is important to know what current error can still be reduced by the next feasible update, motivating work on delayed labels and partial-update budgets together.
Critical weaknesses
  1. Limited technical novelty and theoretical justification: Extracting anchored block gradients amounts to a standard projection onto a parameter subspace, and the bounds rely on classical tools such as uniform convergence and union bounds. I see little technical novelty in these components. Deriving an upper bound on the error does not, by itself, establish the algorithm’s effectiveness. Its benefits need to be substantiated through comparisons with existing methods. In addition, given the streaming setting, I would encourage a theoretical analysis of time and memory complexities.

  2. Insufficient analysis and empirical support for the anchored reachable family: The paper provides insufficient comparative analysis of the benefits of the anchored reachable family over existing methods, or of the bias when the structure retained from the deployed anchor is inaccurate. For example, Theorem 2 characterizes the localization bias, but the paper provides too little quantitative and experimental discussion of how large this bias is relative to the benefits of the anchored reachable family, or when it is small enough to be negligible. Therefore, it does not adequately explain how this structural restriction translates into performance differences relative to existing methods. In the empirical comparison, the experiments show substantial gains over baselines on only one dataset, which I find insufficient evidence for the empirical benefits of the proposed structure.

Questions for authors
  1. What is the main technical contribution beyond classical tools (e.g, standard projection, uniform convergence, and union-bound arguments), and how do the bounds establish a benefit over existing methods?

  2. Under what conditions does the anchored reachable family offer an advantage over existing methods, and how large is the bias when the structure retained from the deployed anchor is inaccurate?

1 comment

Dear reviewer,

Have the authors address your concerns in their replies?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.