acceptodds
Under review as a conference paper at ICLR 2027

What Makes an Update Harmful? The Dynamics of Functional Risk in Model Adaptation

Abstract

What makes an adaptation update harmful to what a model already knows? Existing preservation methods constrain updates using historical gradients, activation-derived directions, or protected subspaces. We take a more direct functional view: for a protected loss, we ask how a concrete update proposed by the optimizer changes it. To first order, this change is the inner product between the update and the gradient of the protected loss, which we call the functional risk of the update. Estimating this gradient from finite reference data raises two questions for any method built on such estimates: how accurately must the signal be estimated to judge an update, and how long does it remain informative as the model changes? Judging an update requires only the gradient component along it, but AdamW proposals are nearly orthogonal to the protected gradient, and we prove that any procedure then needs a number of reference examples that grows with the inverse squared cosine between the two. Correcting an update is easier: a correction along the estimated gradient lowers the protected loss whenever the estimate is aligned with the true gradient. Across model states, a cached estimate changes later decisions, most strongly after a task switch. Motivated by these findings, we introduce Dynamic Functional Risk Control (DFRC), a lightweight controller that periodically refreshes its estimates and minimally modifies candidate updates whose predicted risk exceeds a tolerance; its forgetting over a run is bounded independently of detection accuracy. In sequential adaptation of two model families over three task orders, DFRC improves average accuracy and reduces forgetting relative to sequential LoRA, and further reduces forgetting when combined with an existing LoRA continual-learning method.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
model adaptation, continual learning, catastrophic forgetting, parameter-efficient fine-tuning, capability preservation, large language 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_RHxsOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper defines an update’s “functional risk” as the inner product between the protected-task loss gradient and the optimizer’s proposed parameter update, which approximates the change in that loss. The paper separates prediction error into sampling noise, stale gradient estimates, and higher-order effects. Its central insight is that detecting harmful updates can be difficult while reducing their damage can be easier: the proposed updates are nearly orthogonal to protected gradients, making their effects difficult to estimate, but a correction along a sufficiently aligned gradient estimate can still help. The authors introduce Dynamic Functional Risk Control (DFRC), which periodically refreshes protected gradients and minimally adjusts updates that violate estimated constraints. Across two models and three orders of four classification tasks, DFRC-MC improves average accuracy by 1.06 percentage points and reduces forgetting by 1.45 points relative to sequential LoRA, with additional retention benefits when combined with MBF.

Critical strengths

The paper provides a clear explanation of why noisy estimates can support useful corrections even when they cannot reliably classify individual updates as harmful. Assessing the actual optimizer proposal is well motivated because AdamW’s momentum and adaptive scaling can change the effect implied by the raw gradient. The mechanism studies strengthen the argument through independent reference and evaluation sets, comparisons with larger reference sets, and equal-norm random-direction controls, which support the importance of the protected-gradient direction. The authors also report counterexamples and failed hypotheses transparently. Finally, DFRC improves retention without materially reducing current-task acquisition in the reported experiments, and its compatibility with MBF suggests practical value as an additional controller.

Critical weaknesses

I think your conclusion describes the weaknesses the best, particularly the "DFRC’s 256-example memory has not been compared with replay of the same examples"

Questions for authors

The paper notes that DFRC’s 256-example memory has not been compared with replay of the same examples. I would be interested in seeing this comparison if results become available during the rebuttal period.

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

Error of estimating a functional-risk score from a finite sample is decomposed into a sampling term, a staleness term, and a higher-order remainder. The paper finds that for the Adam optimizer, the sampling term dominates the others. While deciding whether an update is harmful requires a number of samples that grows with the inverse squared cosine between the update and the protected gradient, correcting it is much easier. From this motivation, the paper suggests Dynamic Functional Risk Control (DFRC), a method that applies the smallest correction that satisfies the estimated constraint.

Critical strengths

The paper provides comprehensive theoretical analysis on 1) each sampling term, a staleness term, and a higher-order remainder, and 2) detecting the error is difficult.

Critical weaknesses

Novelty related

  • Trivial decomposition: The decomposition of the error into a sampling term, a staleness term, and a higher-order remainder seems to be a trivial result drawn from the first-order approximation and the decomposition into the gradient at the current point θt\theta_t and the gradient estimate at the past point θs\theta_s.
  • Naive combination of GEM (A-GEM) with the Adam optimizer: DFRC seems to apply GEM (A-GEM) techniques to Adam updates (while refreshing the protected gradients), limiting the novelty of the proposed method.

Experiments related

  • Limited baseline: As demonstrated in the related work section, optimization techniques for model adaptation have been proposed for a long time (EWC, GEM, A-GEM, OGD). However, the main experimental results in Table 1 do not provide comparisons with these methods (with and without MBF).
  • Cost comparison: The method requires computing the smallest correction that satisfies the estimated constraint for each task. As this can take substantial time, it is better to report its cost to show that the comparison in Table 1 is fair.

Presentation related

  • Connection between Sections 3 and 4: The paper provides theoretical analysis on 1) each decomposed error term, and 2) detecting the error is difficult, while correcting it is easy. However, it was hard to follow how the analysis motivates DFRC.
Questions for authors
  • Would you provide a comparison with other optimization methods (EWC, GEM, A-GEM, OGD) with and without MBF?
Reviewer_cfRkOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper investigates catastrophic forgetting during sequential language-model adaptation by analyzng how individual optimizer updates affect previously learned tasks. The authors introduce a functional-risk measure based on the alignment between an update and the gradient of a protected task. They derive theoretial bounds on the reliability of this measure and propose Dynamic Functional Risk Control (DFRC), which modifies potentially harmful updates through gradient projection. Experiments on language models show reduced forgetting during sequential LoRA adaptation.

Critical strengths

The main weakness is the limited conceptual advance toward solving the underlying continual-learning problem. Rather than addressing why learned knowledge is overwritten or how it could be preserved through better represntations or learning mechanisms, the paper focuses on detecting and correcting potentially harmful parameter updates. While mathematically well motivated, this approach primarily mitigates the symptoms of catastrophic forgetting.

Furthermore, the proposed method is closely related to established gradient-projection approaches such as GEM and A-GEM, making the algorithmic novelty relatively limited. The empirical results are also not entirely convincing: standalone DFRC performs worse than the MBF baseline in terms of forgetting, and the experiments use only a single random seed. Importanly, the authors do not compare against experience replay using the same stored examples, leaving the practical advantage of their approach unclear.

Critical weaknesses
  • The paper provides a rigorous mathematical analysis of how optimizer updates affect previosly learned tasks.
  • The distinction between the difficulty of detecting harmful updates and the relative ease of correcting them is an interesting theoretical insight.
  • The authors evaluate their approach on reasonably large language models and demonstrate mesurable reductions in forgetting.
  • The method can be combined with existing continual-learning approaches, suggesting some practical flexibility.
Questions for authors

no questions

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

This paper presents series of theoretical analysis for detecting and correcting harmful updates (which introduces forgetting) during continual learning. The paper considers a method where protected gradients are periodically estimated from subset of dataset and leverage this information to detect or correct harmful updates. The theory analysis three main concerns: sampling error, staleness error, and higher-order effects. Both theoretically and empirically the paper shows that detecting is hard but correcting is easy for harmful updates. Motivated by these theoretical results, the paper proposes DFRC, which is a novel method that prevents harmful updates and thereby alleviates forgetting learned knowledge. DFRC corrects the update only when it does not fulfill the protect condition, and otherwise keep it same. DFRC is evaluated on sequential adaptation of large language models, and when combined with MBF, surpasses baseline methods.

Critical strengths

The theoretical analysis is useful and considers important and interesting problem. The theory is build upon practical constrains and assumptions (e.g. periodically computing reference gradient for practical computational cost), and well studies potential estimation error (e.g. sampling error). The theoretical results are also well validated in the LLM experiments. However I did not carefully check the correctness of the theory in Appendix.

Critical weaknesses

Effectiveness of the proposed DFRS method is questionable.

  1. Insufficient baselines
  • The comparison lacks important baselines in the field, such as EWC [1], LwF [2], A-GEM [3], O-LORA [4], S-LORA [5].
  • I understand huge computation of LLM experiments, but I think at least the latest and the strongest baselines (O-LORA and S-LORA) should be included to increase my score.
  1. Improvement is marginal and standard deviation is not reported
  • In the main result in Table 1, the improvement of the proposed method compared to existing baselines are quite marginal (less than 0.01), while the standard deviation is not reported. Therefore it is hard to check whether the proposed method is effective. Also combining DFRC to MBF improves performance, but when only applied DFRC the performance is worse than MBF.

References

[1] Kirkpatrick, James, et al. "Overcoming catastrophic forgetting in neural networks." Proceedings of the national academy of sciences 114.13 (2017): 3521-3526.

[2] Li, Zhizhong, and Derek Hoiem. "Learning without forgetting." IEEE transactions on pattern analysis and machine intelligence 40.12 (2017): 2935-2947.

[3] Chaudhry, Arslan, et al. "Efficient lifelong learning with a-gem." arXiv preprint arXiv:1812.00420 (2018).

[4] Wang, Xiao, et al. "Orthogonal subspace learning for language model continual learning." Findings of the Association for Computational Linguistics: EMNLP 2023. 2023.

[5] Yang, Lina, et al. "SLoRA: Balancing Plasticity and Forgetting in Large Language Models for Continual Learning." Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Questions for authors

Q1. Could you include O-LORA and S-LORA as baselines? Q2. Could you report std in Table 1?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.