acceptodds
Under review as a conference paper at ICLR 2027

Sparse Autoencoders Can Induce Local Geometry for Mitigating Catastrophic Forgetting

Abstract

Catastrophic forgetting remains a challenge in supervised continual learning, as adapting to new tasks causes representation drift and loss of previously learned information. This raises the question of which representational changes should be constrained while preserving plasticity. We therefore introduce Sparse Autoencoder Geometric Anchoring (SAGA), which converts a Sparse Autoencoder (SAE) trained at each task boundary into local, low-rank constraints on representation drift. Replay examples specify where the constraints apply, while leading right singular directions of the encoder Jacobian specify which changes are penalized. The resulting geometry is input dependent and can be retained without keeping the SAE for subsequent tasks' training. We establish its local relation to latent code matching, define its spectral truncation error, and derive a conditional bound on frozen-readout drift. With replay and snapshot history, the method achieves the highest mean Class-IL accuracy and lowest mean forgetting among the evaluated ER, DER++, ER-ACE, and SGD baselines on three pretrained ViT-B/16 benchmarks, with accuracy gains of 12.15, 4.79, and 3.22 percentage points over the strongest conventional baseline on CIFAR-100, CUB-200, and ImageNet-R, respectively. From scratch results are more mixed, while rank ablations and geometry controls show that SAE-induced low-rank geometry offers useful stability-plasticity trade-offs.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual Learning, Sparse Autoencoders, Catastrophic Forgetting, Jacobian

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

This paper proposes Sparse Autoencoder Geometric Anchoring (SAGA), a representation-level regularization method for mitigating catastrophic forgetting in class-incremental learning. The key idea is to use Sparse Autoencoders (SAEs) to identify input-dependent, task-specific directions in feature space that should be protected during subsequent training. Specifically, SAGA computes the Jacobian of a task-boundary SAE encoder and applies singular value decomposition to construct low-rank geometric constraints on representation drift. These constraints are combined with experience replay and historical snapshots to balance stability and plasticity.

The paper provides theoretical analyses connecting Jacobian-based anchoring to latent feature matching, characterizing low-rank approximation errors, and establishing a conditional bound on old-task output drift. Experiments are conducted on CIFAR-100, CUB-200, ImageNet-R, and TinyImageNet using pretrained ViT-B/16 and randomly initialized ResNet-18 backbones. Results show that SAGA outperforms the evaluated conventional continual learning baselines across three pretrained settings, while results on from-scratch training are more mixed. Additional ablations investigate the effects of replay, historical anchoring, retained rank, and alternative geometric constraints.

Critical strengths

1. An interesting perspective on representation-level regularization.

The paper explores the use of SAE encoder Jacobians to construct sample-dependent, low-rank constraints on representation drift. This provides a potentially useful perspective on selectively preserving representations during continual learning, rather than uniformly penalizing feature changes.

2. Empirical evaluation across different training settings.

The paper evaluates the proposed method on both pretrained and randomly initialized backbones and includes ablations on replay, historical anchoring, and alternative geometric constraints. The results show improvements over several conventional continual learning baselines in pretrained settings, while also revealing limitations in other configurations.

Critical weaknesses
  1. The paper primarily compares SAGA against relatively old continual learning baselines, such as ER, DER++, and ER-ACE, while omitting comparisons with more recent approaches, particularly PEFT/LoRA-based methods for pretrained models. This makes it difficult to assess the competitiveness of SAGA against current continual learning techniques. More importantly, SAGA retains additional historical snapshots containing images, representations, and geometric factors, whereas the conventional baselines operate with a fixed replay buffer. Without comparisons under matched total memory budgets, it is unclear whether the reported improvements primarily result from the proposed geometric constraints or the additional memory resources.

  2. The central claim of the paper is that SAE-induced local geometry provides more effective constraints for mitigating forgetting. However, the experiments do not convincingly establish this advantage. First, Jacobian-based and latent-based anchoring employ different loss normalization schemes, making their comparisons potentially unfair. Second, SAE-Jacobian anchoring does not consistently outperform simpler alternatives such as PCA and classifier SVD, raising questions about the necessity of introducing SAEs. Furthermore, the paper lacks direct visualizations or empirical analyses demonstrating that the identified geometric directions correspond to task-relevant information or genuinely help preserve previously learned knowledge. Consequently, the effectiveness and necessity of the proposed geometric mechanism remain insufficiently validated.

Questions for authors
  1. Given that the performance improvements appear to rely heavily on experience replay, could the authors provide a baseline that retains the same replay strategy and historical snapshots but removes the Jacobian regularization loss to isolate the actual contribution of the proposed geometric constraints?

  2. Since SAE-Jacobian anchoring does not consistently outperform simpler alternatives such as PCA and classifier SVD, could the authors further demonstrate the necessity of SAE-induced geometry and provide comparisons with latent anchoring under matched regularization strengths?

4 comments

We thank the reviewer for all the time and effort spent on providing us with helpful comments!

We thank the reviewer for finding SAGA’s use of SAE Jacobians for sample-dependent, low-rank representation anchoring an interesting perspective on continual learning. We also appreciate their recognition of the evaluation across pretrained and randomly initialized backbones, the ablations, and the balanced reporting of both gains and limitations.

We did our best to address all specific comments and provide additional experimental analysis. We address each comment in a separate comment. We hope that these revisions satisfactorily address the reviewer’s concerns!

We will be happy to answer if any of the issues remain.

AuthorsOpenReview ↗

Baseline with the same replay strategy and snapshots (Q1)

Thanks for this remark. We now report storage for every method and separate the effect of the regularizer from the effect of data access.

Storage. SAGA can store the per-sample geometry, but it can equivalently store each snapshot's SAE and the anchored images' features, and recompute the directions at the task boundary (SAE takes about 3 MiB per snapshot), still being an explicit, rank-controllable quadratic formulation of SAE-based preservation. With the second implementation, SAGA retains 193–229 MiB across our scenarios: the snapshot images plus 27 MiB of features and SAEs. We will clarify this option in the manuscript.

  1. Same data and training, no anchor. We keep SAGA's snapshots, replay sampling and schedule fixed and set only the anchor weight to 0. Accuracy drops by 12.9 / 10.0 / 11.5 points (CIFAR-100-224 / CUB / ImageNet-R), and by 7.7 on ResNet-18 CIFAR-100. Therefore, the gain comes from our Jacobian-based regularizer, not from greater access to historical data or additional training passes.
  2. Equal storage. We give ER, DER++ and ER-ACE SAGA's exact stored images (1,156–1,472), replayed on SAGA's schedule, with each baseline's best of 3 learning rates. SAGA still leads the best baseline by +13.0 on CIFAR-100-224 and +2.5 on CUB, and trails ER-ACE by 5.0 on ImageNet-R. These results suggest that even under equal storage in some cases, SAGA’s anchoring can yield substantial improvements.
  3. Compute. The 6.4–7.1× figure was relative to SGD. Against ER and ER-ACE on the same GPU, SAGA is only about 1.5× slower per task.
MethodStoredCIFAR-100-224CUBImageNet-R
Best baseline, 500 images72 MiB63.8668.8958.36
Best baseline, SAGA's images166–211 MiB62.9971.1866.60
SAGA, without anchor193–229 MiB63.1363.6950.06
SAGA193–229 MiB76.0173.6861.58

Overall, the results show that the geometric anchor itself (not the additional memory) is the main source of SAGA's gains: on identical stored data it adds 10–13 points on the pretrained benchmarks. We will add these storage-matched comparisons and the memory accounting to the revised paper.

AuthorsOpenReview ↗

Matching loss normalization schemes and further comparison with geometric alternatives

We thank the reviewer for raising comments about matching loss normalization schemes and further comparisons with geometric alternatives to show Jacobian benefits.

Fair latent comparison

We have now matched the normalization and independently tuned the regularization coefficients for Jacobian, latent, and frozen-support anchoring for a fair comparison.

The results are consistent with our theoretical comparison: at fixed support and full retained rank, Jacobian anchoring is the local quadratic approximation of latent matching, with the equal inside a fixed support at r = K. They show broadly comparable performance, with latent matching performing better in some settings. The latent loss is discontinuous: 22–45% of it has zero gradient once latents leave the TopK. The Jacobian loss is a smooth quadratic with no such dead zones. Therefore, the formulations differ when the representation moves across TopK support boundaries: latent matching can lose its restoring gradient for snapshot latents that are no longer active, whereas the Jacobian anchor continues to constrain the snapshot directions. This makes it less sensitive to the anchor weight, although final accuracy at tuned weights is the same. We will revise the paper to avoid claiming a general accuracy advantage for Jacobian anchoring, and to highlight this Jacobian-based anchoring characteristic.

Table 1. Jacobian vs latent vs frozen-support latent at matched strength. One normalization for all three objectives; each tuned independently on the same lr × anchor-weight grid on the validation split; class-IL test accuracy, mean ± std over 3 seeds.

SettingKJacobianLatentLatent, frozen support
CUB-200 (ViT)1673.33 ± 0.6073.85 ± 0.5473.25 ± 0.57
CUB-200 (ViT)6474.56 ± 1.0474.72 ± 0.9774.71 ± 0.60
ImageNet-R (ViT)1662.32 ± 0.9964.15 ± 1.4961.10 ± 0.38
ImageNet-R (ViT)6464.58 ± 0.7166.19 ± 1.1264.40 ± 0.31
CIFAR-100 (ResNet-18)1627.81 ± 2.8527.26 ± 2.1226.53 ± 0.88
TinyImageNet (ResNet-18)1635.99 ± 1.3032.17 ± 5.09 †32.23 ± 6.25

This does not undermine our contribution, which is still an explicit, rank-controllable quadratic formulation of SAE-based preservation. This enables explicit rank control, which latents cannot do. In matched-strength experiments, accuracy remains relatively high even at low ranks. This supports rank truncation as an interpretable way to control how many representation directions are protected, while retaining flexibility in the others.

Table 2. Sensitivity to the anchor weight

Anchor typeAccuracy lost using one shared weight (mean / worst setting)Drop one grid step from own best weight
Jacobian0.27 / 0.831.06
Latent0.68 / 2.101.58
Latent, frozen support0.05 / 0.281.24

Continuation regarding the geometry comparisons in the next comment.

AuthorsOpenReview ↗

... further comparison with geometric alternatives

Geometrical comparisons

To more thoroughly examine what benefits can SAE Jacobian-based anchoring bring, we analyze TopK support changes, centered kernel alignment (CKA), drift outside of the protected subspace and leakage. Compared with PCA and classifier-SVD anchoring, SAGA achieves comparable global stability by CKA while consistently obtaining lower drift outside the protected subspace on both CUB-200 and CIFAR-100-224 than other methods. It also best preserves local structure, producing the fewest TopK support changes in both settings. SAGA and PCA have a similar leakage, and SVD, as expected, obtains a lower value. Even though SAGA does not have the lowest readout leakage, it achieves the lowest forgetting on CUB-200 and competitive forgetting on CIFAR-100-224. This highlights an interesting distinction in our analysis: leakage measures how much of the classifier’s readout lies outside the protected subspace, but not necessarily how much continual learning updates actually move representations in those specific directions. Thus, static classifier sensitivity should be used along with other measures.

Overall, these findings support SAGA as preserving global representation stability and local SAE support structure. We will clarify that leakage is a conditional sensitivity factor in our bound, not a standalone predictor of forgetting. We will also add the additional analyses to the appendix of the paper and mention it shortly in the main paper. Moreover, this distinction motivates a direction for future work: jointly tracking readout sensitivity and update-induced representation drift may help identify which task-relevant information is disrupted first during continual learning, providing a more mechanistic account of forgetting – we will also mention it in the updated manuscript.

Table: Earlier tasks at the last task boundary. Stability and forgetting of tasks 1–9 after training on task 10, by anchor geometry (r = 16, normalized weights). Mean over earlier tasks, then over 3 seeds.

DatasetGeometryLeakage βLeakage LCKA ↑TopK change ↓Drift outside subspace ↓Forgetting, live head ↓Forgetting, frozen head ↓
CUB-200SAE (SAGA)0.6190.9630.9660.14748.814.22.2
CUB-200PCA0.6120.9350.9670.19954.616.83.2
CUB-200Classifier SVD0.5860.8500.9240.25563.521.65.6
CIFAR-100-224SAE (SAGA)0.5870.9370.9640.09744.424.81.3
CIFAR-100-224PCA0.5670.8570.9650.17449.324.21.0
CIFAR-100-224Classifier SVD0.5450.7850.9410.14650.022.43.9

Interpretation of the symbols: Leakage β=∥A(I−P)∥op\beta = \lVert A(I - P) \rVert_{\mathrm{op}} and L=∥A(I−P)∥F2 / ∥A∥F2L = \lVert A(I - P) \rVert_F^2 \,/\, \lVert A \rVert_F^2 measure how much of the task's readout A lies outside the protected subspace P (computed at the task's own boundary; for the SAE geometry, averaged over images). CKA: similarity of the task's held-out features now vs. at its own boundary. TopK change: share of an image's active SAE latents that changed since its boundary. Drift outside subspace: ∥(I−Pi) (hnew−hold)∥\lVert (I - P_i)\,(h_{\text{new}} - h_{\text{old}}) \rVert, the feature change outside the image's own protected directions. Forgetting: drop in a task's accuracy from its own boundary to the last one, with the current classifier (live) or with the classifier saved at that task's boundary (frozen).

Reviewer_bjjTOpenReview ↗
Rating
3
Confidence
2
Informativeness
3(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Sparse Autoencoder Geometric Anchoring (SAGA) for supervised continual learning. At each task boundary, a TopK sparse autoencoder is trained on representations from a reservoir-buffer snapshot. The method extracts leading right-singular directions and singular values of the encoder Jacobian at each stored example, using these factors to penalize subsequent representation drift. The SAE itself can then be discarded. The full method combines this regularizer with supervised replay and historical snapshots. The paper analyzes fixed-support equivalence to latent matching, spectral truncation, and a conditional bound on frozen-readout drift. Experiments cover three pretrained ViT-B/16 settings and two ResNet-18 settings trained from scratch. SAGA outperforms the evaluated conventional baselines on the pretrained benchmarks, while results from scratch and comparisons against alternative geometries are mixed.

Critical strengths
  1. A concrete and interesting use of SAE geometry. Exporting input-dependent sensitivity factors provides a clear alternative to retaining an SAE and matching its nonlinear codes. The distinction between where representations are anchored and which directions are constrained is useful.
  2. Meaningful empirical results with pretrained backbones. Table 3 reports Class-IL accuracy improvements of 12.15, 4.79, and 3.22 percentage points over the strongest evaluated conventional baseline on CIFAR-100, CUB-200, and ImageNet-R, respectively. These results motivate further investigation of geometric anchoring.
  3. Careful qualification of theoretical and empirical claims. The paper explicitly distinguishes geometric energy from implemented loss normalization, acknowledges support changes in hard TopK encoders, and states that the readout bound is conditional. It also reports unfavorable results and discloses the growing memory requirements of History.
  4. Relevant ablations. Replay/history combinations, retained rank, alternative geometries, and dictionary size are examined. The normalized geometry comparisons are particularly valuable because they test whether the benefit is specific to SAE-derived constraints.
Critical weaknesses
  1. The contribution of geometric anchoring is insufficiently separated from historical memory and replay protocol. Section 4 explains that History stores complete reservoir snapshots, representations, and anchor factors from all completed task boundaries. Consequently, Table 3 compares methods with different total memory requirements. The ablations also lack a replay-only control using the same retained snapshot collection and a comparable sampling/optimization budget. This does not invalidate the reported full-system results. However, the substantial improvement from adding History in Tables 1–2 cannot be attributed cleanly to preserving historical geometric constraints rather than the broader historical data coverage and additional training passes. This is important because the headline improvements concern the full configuration.
  2. The latent comparison does not isolate the proposed mechanism. Equations 6–7 use different reductions: the Jacobian loss divides geometric energy by the retained rank, while latent matching sums latent-coordinate errors. Within a fixed support cell and at full retained rank, the underlying energies are identical, but the implemented losses differ by a factor of r. Section 3 explicitly states that the comparison is not matched in regularization strength. Accordingly, the four-out-of-five accuracy advantage over latent matching does not establish the benefit of replacing nonlinear code matching with frozen local geometry. A scale-matched comparison, with clearly documented validation-based coefficient selection for both objectives, is needed to assess this central design choice.
  3. The distinctive benefit of input-dependent SAE geometry remains only partially established. Under the controlled protocol in Table 6, SAE-Jacobian anchoring is essentially tied with PCA on ImageNet-R (61.44 versus 61.37), is below classifier SVD on CIFAR-100 (76.08 versus 77.34), and has a modest advantage over PCA on CUB-200 (72.55 versus 71.64). Table 7 also shows mixed accuracy/forgetting trade-offs. These results support SAE geometry as a viable option, as the authors appropriately state, but provide limited evidence for why sample-dependent SAE geometry is preferable to simpler constructions. Equation 12 identifies the relevant unprotected readout sensitivity through β, yet this quantity is not evaluated. Moreover, the existing controls change locality, construction, and spectral weighting together, so they do not isolate the value of locality itself.
  4. The practical accuracy–resource trade-off is not sufficiently characterized. Appendix E reports a 6.4–7.1× training-time overhead relative to SGD, but does not report timing against replay baselines or the global geometry alternatives. Appendix D.4 further indicates that SAE factors are stored per example, whereas the reference bases are shared across examples. Thus, discarding the SAE does not by itself establish a storage advantage. Reporting total retained bytes and runtime for the relevant alternatives would materially clarify the practical value of the method.
  5. The baseline coverage is limited relative to the proposed contribution. Tables 3–4 compare SAGA primarily against ER, DER++, and ER-ACE, with SGD serving as an unregularized reference. These comparisons establish performance relative to conventional replay methods, but do not adequately position SAGA against representation-preservation or subspace-based continual learning approaches. The related work discusses methods such as Backward Feature Projection, PODNet, and GPM, yet none is evaluated directly. The random, PCA, and classifier-SVD controls are useful for comparing geometries within the authors’ framework, but do not replace comparisons against established methods addressing closely related objectives. A representative comparison under a compatible protocol, with total memory and tuning budgets documented, would help establish the added value of SAE-derived local geometry.
Questions for authors
  1. What happens under matched historical-data access? Can you provide a replay-only control that uses the same historical snapshot collection and comparable optimization budget? A persistent gain from geometric anchoring would substantially strengthen the interpretation of Tables 1–3.
  2. Were latent and Jacobian coefficients independently selected? Please specify the search ranges and whether they account for the factor-r difference. A directly scale-matched comparison would affect my assessment of the Jacobian-versus-latent contribution.
  3. Does the learned local geometry preserve functionally relevant directions? Can an analysis of existing checkpoints compare SAE, PCA, and classifier geometries using normalized readout leakage or observed frozen-readout drift? Evidence connecting the geometry to retained function would strengthen the mechanistic contribution.
  4. What are the actual resource requirements? Please report total snapshot/image/factor storage and training time against ER/DER++ and at least one global geometry control. Please also clarify the validation and tuning protocol used for these controls and the conventional baselines.
8 comments

We would like to thank the reviewer for all the time and effort spent on providing us with helpful comments!

We thank the reviewer for noticing the strenghts of our paper:

  • SAGA’s use of SAE geometry,
  • promising results with pretrained backbones,
  • careful qualification of our claims and limitations,
  • breadth of the ablations.

We did our best to address all specific comments and provide requested additional experimental analysis. We address each comment in a separate comment. We hope that these revisions satisfactorily address the reviewer’s concerns.

We will be more than happy to adress any remaining issues, if such occur.

AuthorsOpenReview ↗

Do SAGA’s gains hold under matched memory and compute budgets? (Q1, W1)

Thanks for this comment! We agree that the original comparison was not matched by historical data. We now report storage for every method and separate the effect of the regularizer from the effect of data access.

SAGA can store the per-sample geometry, but it can equivalently store each snapshot's SAE and the anchored images' features, and recompute the directions at the task boundary (SAE takes about 3 MiB per snapshot), still being an explicit, rank-controllable quadratic formulation of SAE-based preservation. With the second implementation, SAGA retains 193–229 MiB across our scenarios: the snapshot images plus 27 MiB of features and SAEs. We will clarify this option in the manuscript.

Below, we discuss specific aspects regarding historical and memory-matched baselines.

  1. Same data and training, no anchor. We keep SAGA's snapshots, replay sampling and schedule fixed and set only the anchor weight to 0. Accuracy drops by 12.9 / 10.0 / 11.5 points (CIFAR-100-224 / CUB / ImageNet-R), and by 7.7 on ResNet-18 CIFAR-100. Therefore, the gain comes from our Jacobian-based regularizer, not from greater access to historical data or additional training passes.
  2. Equal storage. We give ER, DER++ and ER-ACE SAGA's exact stored images (1,156–1,472), replayed on SAGA's schedule, with each baseline's best of 3 learning rates. SAGA still leads the best baseline by +13.0 on CIFAR-100-224 and +2.5 on CUB, and trails ER-ACE by 5.0 on ImageNet-R. These results suggest that even under equal storage, SAGA’s anchoring can yield substantial improvements.
  3. Compute. The 6.4–7.1× figure was relative to SGD. Against ER and ER-ACE on the same GPU, SAGA is only about 1.5× slower per task.
MethodStoredCIFAR-100-224CUBImageNet-R
Best baseline, 500 images72 MiB63.8668.8958.36
Best baseline, SAGA's images166–211 MiB62.9971.1866.60
SAGA, without anchor193–229 MiB63.1363.6950.06
SAGA193–229 MiB76.0173.6861.58

Overall, the results show that the geometric anchor itself (not the additional memory) is the main source of SAGA's gains: on identical stored data it adds 10–13 points on the pretrained benchmarks. We will add these storage-matched comparisons and the memory accounting to the revised paper.

AuthorsOpenReview ↗

Does SAGA’s geometry help beyond penalty scale and regularization strength? (Q2, W3)

We thank the reviewer for noting that our original latent comparison does not isolate the proposed mechanism, which we also stated in the original manuscript “the differing loss reductions in Eqs. 6 and 7 prevent attributing every difference solely to the choice of local geometry.” We have now performed additional experiments that independently tuned the regularization coefficients for Jacobian, latent, and frozen-support anchoring for a fair comparison.

The results align with our theoretical comparison. At fixed support and full retained rank, Jacobian anchoring is the local quadratic approximation of latent matching, with the equal inside a fixed support at r = K. They show broadly comparable performance, with latent matching performing better in some settings. The latent loss is discontinuous: 22–45% of it has zero gradient once latents leave the TopK. The Jacobian loss is a smooth quadratic with no such dead zones. Therefore, the formulations differ when the representation moves across TopK support boundaries: latent matching can lose its restoring gradient for snapshot latents that are no longer active, whereas the Jacobian anchor continues to constrain the snapshot directions. This makes it less sensitive to the anchor weight, although final accuracy at tuned weights is the same. We will revise the paper to avoid claiming a general accuracy advantage for Jacobian anchoring, and to highlight this Jacobian-based anchoring characteristic.

Table 1. Jacobian vs latent vs frozen-support latent at matched strength. One normalization for all three objectives; each tuned independently on the same lr × anchor-weight grid on the validation split; class-IL test accuracy, mean ± std over 3 seeds.

SettingKJacobianLatentLatent, frozen support
CUB-200 (ViT)1673.33 ± 0.6073.85 ± 0.5473.25 ± 0.57
CUB-200 (ViT)6474.56 ± 1.0474.72 ± 0.9774.71 ± 0.60
ImageNet-R (ViT)1662.32 ± 0.9964.15 ± 1.4961.10 ± 0.38
ImageNet-R (ViT)6464.58 ± 0.7166.19 ± 1.1264.40 ± 0.31
CIFAR-100 (ResNet-18)1627.81 ± 2.8527.26 ± 2.1226.53 ± 0.88
TinyImageNet (ResNet-18)1635.99 ± 1.3032.17 ± 5.09 †32.23 ± 6.25

This does not undermine our contribution, which is still an explicit, rank-controllable quadratic formulation of SAE-based preservation. This enables explicit rank control, which latents cannot do. In matched-strength experiments, accuracy remains relatively high even at low ranks. This supports rank truncation as an interpretable way to control how many representation directions are protected, while retaining flexibility in the others.

Table 2. Sensitivity to the anchor weight

Anchor typeAccuracy lost using one shared weight (mean / worst setting)Drop one grid step from own best weight
Jacobian0.27 / 0.831.06
Latent0.68 / 2.101.58
Latent, frozen support0.05 / 0.281.24

As for a search ranges, in the additional analysis shown here, we used the search space defined below:

BackboneLearning ratesAnchor weights λK
ViT-B/16 (pretrained, Adam){1e-5, 3e-5}{0.1, 0.3, 1, 3, 10, 30}16, 32, 64
ResNet-18 (scratch, SGD)0.03, 0.1{0.1, 0.3, 1, 3, 10, 30}16, 32, 64

Additionally, for each K we tuned ranks (r) from the following list: [4, 8, 16, 32, 64] where r < K. We will add these additional analyses to the paper.

Reviewer_md79OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Sparse Autoencoder Geometric Anchoring (SAGA) that takes as an input task-boundary SAE encoder Jacobians and converts them to low-rank penalizations on representation drift. It is combined with replay mechanism to improve pre-trained models in continual learning setup.

Critical strengths

I commend Authors for very nice insights on the anchor location of representations vs in which directions they are constrained. Easy to follow and useful. I like the analysis of TopK support changes and sensitivity outside of the protected subspace. Experiments report useful findings, such as rank, geometry, reply, and ablations with separate validation tuning.

Critical weaknesses

I found the following four critical weaknesses in the reviewed paper:

W1: The main comparison is limited by resource differences. The proposed method stores complete buffer snapshots, representations, and per-sample geometry at every task boundary, while the reference methods retain only fixed buffers. This confounds the reported gains between the proposed regularizer and significantly greater access to historical data and additional training phases. The acknowledged 6.4–7.1× runtime relative to SGD does not establish competitiveness against replay methods under matched memory or compute.

W2: Difficult to actually measure the direct impact of SAE-induced local geometry. Reading Table 6, I can see that SAE geometry surpasses PCA geometry only by 0.07 points on ImageNet-R, while classifier SVD performs better on CIFAR-100. CUB-200 advantage over PCA is modest relative to the reported variability. These controls are valuable, but stronger evidence is needed that sample-specific SAE geometry justifies its additional complexity.

W3: Latent comparison mixes geometry and regularization strength. While the paper recognizes the mismatch between Jacobian loss divided by ranks and latent matching sums coordinate errors, there is no experimental analysis of this issue. There is a need to clearly establish whether benefits arise from locality, support handling, compression, or simply a different penalty scale.

W4: The theory is disjoint from empirical study. The proposed frozen bound is interesting, but depends on the unmeasurable leakage term. It says nothing about if SAE-based geometry is superior. Generalization capabilities also remain vague: all experiments use ten tasks and a single class order, and SAGA underperforms stronger replay baselines in final accuracy when trained from scratch.

Questions for authors

I would like Authors to address these questions and other issues raised in Weakness section.

Q1: Do the gains persist when ER/DER++ receive equivalent total storage and training compute, or when SAGA’s historical snapshots are constrained to a fixed total budget?

Q2: How do Jacobian anchoring, latent matching, frozen-support matching, and full-feature distillation compare after matching loss scale and independently tuning regularization strength?

Q3: Does assigning each example another example’s SAE geometry reduce performance? This would directly test whether sample-specific locality matters beyond the distribution of directions and weights.

Q4: Can you measure or approximate the readout leakage and TopK changes, and test whether they predict forgetting across geometries, longer task sequences, and different class orders?

I look forward to the discussion period.

5 comments

We thank the reviewer for all the time and effort spent on providing us with helpful comments. We would like to thank for recognizing the clarity of our distinction between where representations are anchored and which directions are constrained, as well as the useful analyses of TopK support changes, sensitivity outside the protected subspace, and the rank, geometry, and replay ablations. We did our best to address all specific comments and provide additional experimental analysis. We address each comment in a separate comment. We hope that these revisions satisfactorily address the reviewer’s concerns.

AuthorsOpenReview ↗

Do SAGA’s gains hold under matched memory and compute budgets? (Q1, W1)

Thanks for this remark, we agree that the original comparison was not resource-matched. We now report storage for every method and separate the effect of the regularizer from the effect of data access.

SAGA can store the per-sample geometry, but it can equivalently store each snapshot's SAE and the anchored images' features, and recompute the directions at the task boundary (SAE takes about 3 MiB per snapshot), still being an explicit, rank-controllable quadratic formulation of SAE-based preservation. With the second implementation, SAGA retains 193–229 MiB across our scenarios: the snapshot images plus 27 MiB of features and SAEs. We will clarify this option in the manuscript.

  1. Same data and training, no anchor. We keep SAGA's snapshots, replay sampling and schedule fixed and set only the anchor weight to 0. Accuracy drops by 12.9 / 10.0 / 11.5 points (CIFAR-100-224 / CUB / ImageNet-R), and by 7.7 on ResNet-18 CIFAR-100. Therefore, the gain comes from our Jacobian-based regularizer, not from greater access to historical data or additional training passes.
  2. Equal storage. We give ER, DER++ and ER-ACE SAGA's exact stored images (1,156–1,472), replayed on SAGA's schedule, with each baseline's best of 3 learning rates. SAGA still leads the best baseline by +13.0 on CIFAR-100-224 and +2.5 on CUB, and trails ER-ACE by 5.0 on ImageNet-R. These results suggest that even under equal storage, SAGA’s anchoring can yield substantial improvements.
  3. Compute. The 6.4–7.1× figure was relative to SGD. Against ER and ER-ACE on the same GPU, SAGA is only about 1.5× slower per task.
MethodStoredCIFAR-100-224CUBImageNet-R
Best baseline, 500 images72 MiB63.8668.8958.36
Best baseline, SAGA's images166–211 MiB62.9971.1866.60
SAGA, without anchor193–229 MiB63.1363.6950.06
SAGA193–229 MiB76.0173.6861.58

Overall, the results show that the geometric anchor itself (not the additional memory) is the main source of SAGA's gains: on identical stored data it adds 10–13 points on the pretrained benchmarks. We will add these storage-matched comparisons and the memory accounting to the revised paper.

AuthorsOpenReview ↗

Does SAGA’s geometry help beyond penalty scale and regularization strength? (Q2, W3)

We thank the reviewer for noting that our original Jacobian-latent comparison used different loss normalizations and that our paper recognizes this mismatch, which we stated in the original manuscript “the differing loss reductions in Eqs. 6 and 7 prevent attributing every difference solely to the choice of local geometry.” We have now matched the normalization and independently tuned the regularization coefficients for Jacobian, latent, and frozen-support anchoring for a fair comparison.

The results are consistent with our theoretical comparison: at fixed support and full retained rank, Jacobian anchoring is the local quadratic approximation of latent matching, with the equal inside a fixed support at r = K. They show broadly comparable performance, with latent matching performing better in some settings. The latent loss is discontinuous: 22–45% of it has zero gradient once latents leave the TopK. The Jacobian loss is a smooth quadratic with no such dead zones. Therefore, the formulations differ when the representation moves across TopK support boundaries: latent matching can lose its restoring gradient for snapshot latents that are no longer active, whereas the Jacobian anchor continues to constrain the snapshot directions. This makes it less sensitive to the anchor weight, although final accuracy at tuned weights is the same. We will revise the paper to avoid claiming a general accuracy advantage for Jacobian anchoring, and to highlight this Jacobian-based anchoring characteristic.

Table 1. Jacobian vs latent vs frozen-support latent at matched strength. One normalization for all three objectives; each tuned independently on the same lr × anchor-weight grid on the validation split; class-IL test accuracy, mean ± std over 3 seeds.

SettingKJacobianLatentLatent, frozen support
CUB-200 (ViT)1673.33 ± 0.6073.85 ± 0.5473.25 ± 0.57
CUB-200 (ViT)6474.56 ± 1.0474.72 ± 0.9774.71 ± 0.60
ImageNet-R (ViT)1662.32 ± 0.9964.15 ± 1.4961.10 ± 0.38
ImageNet-R (ViT)6464.58 ± 0.7166.19 ± 1.1264.40 ± 0.31
CIFAR-100 (ResNet-18)1627.81 ± 2.8527.26 ± 2.1226.53 ± 0.88
TinyImageNet (ResNet-18)1635.99 ± 1.3032.17 ± 5.09 †32.23 ± 6.25

This does not undermine our contribution, which is still an explicit, rank-controllable quadratic formulation of SAE-based preservation. This enables explicit rank control, which latents cannot do. In matched-strength experiments, accuracy remains relatively high even at low ranks. This supports rank truncation as an interpretable way to control how many representation directions are protected, while retaining flexibility in the others.

Table 2. Sensitivity to the anchor weight

Anchor typeAccuracy lost using one shared weight (mean / worst setting)Drop one grid step from own best weight
Jacobian0.27 / 0.831.06
Latent0.68 / 2.101.58
Latent, frozen support0.05 / 0.281.24

We will add these additional analyses to the paper.

Comments on the paper

2 comments

We thank all reviewers for their thoughtful comments and constructive feedback. In addition to responding to the individual comments, we have revised the manuscript to incorporate the following results and clarifications:

  • Baseline comparisons (Section 5.2): We added evaluations of ER, DER++, and ER-ACE using SAGA’s exact stored images and replay schedule, reporting accuracy and forgetting over three seeds.
  • Anchoring ablation (Section 5.2): We added a control that preserves SAGA’s snapshots, replay sampling, and training schedule while setting the anchor weight to zero, isolating the contribution of anchoring.
  • Resource accounting (Appendix E): We expanded the storage and runtime analysis, including comparisons with replay baselines, and clarified the distinction between matching historical data access and matching total storage or computation.
  • Class-related structure (Section 5.4 and Appendix F): We added analyses of SAE latent purity, retrieval by Jacobian-subspace overlap, and classification in shared SAE subspaces, together with qualitative visualizations.

We appreciate the opportunity to improve the paper and would be happy to discuss any remaining questions or concerns.

AuthorsOpenReview ↗

summary of the rebuttal

As the rebuttal period comes to an end, we would like to summarize how our clarifications and additional experiments address the concerns raised. We believe they should resolve all the issues identified before rebuttal and significantly reinforce the paper’s strengths noted by the reviewers: its clear geometric perspective, analyses of representation change, meaningful results with pretrained backbones, broad empirical evaluation, and careful reporting.

We have added extensive evidence, including approximately 2,200 GPU hours of experiments on NVIDIA V100 GPUs. We kindly ask the reviewers to consider all these new results in their final assessment. If any concerns remain, we would appreciate specific comments explaining which issues are unresolved and why the new evidence does not address them, so we can further clarify today or tomorrow. If the reviewers are not able to provide such concrete comments, meaning that no critical concerns remain, we kindly ask them to raise their scores.

Summary of the changes:

  • Comparison with Jacobian anchoring at matched regularization strength: Our additional experiments showed comparable performance with Latent-based reference, which aligns with our theoretical analyses and motivation. We further clarified that our contribution lies in rank control and a smooth quadratic constraint, not a universal accuracy advantage over latent matching.

  • Baseline comparison at matching budgets: We added matched ER, DER++, and ER-ACE comparisons, as well as targeted comparisons with recent SAE-based methods. This places SAGA against both established continual-learning references and methods closest to its contribution, addressing the concern that the evaluation lacked relevant comparisons in fair conditions.

  • Baseline coverage: We clarified that the newest reference method used in our experiments (latent matching) was proposed in 2026, and therefore we do use very recent methods. We clarified that this is the method closest to our contribution, and that is why it is the central comparison, however complementary experiments are useful as they place SAGA as competitive against established continual-learning references.

  • Effect of anchoring: The anchor-weight-zero ablation isolated anchoring while holding the rest of the setup fixed. This showed that the SAGA gains come from the proposed geometric constraint and not from the replay or historical snapshots alone.

  • Relevance of protected geometry: Our analyses based on class structure showed that the local subspaces retrieve same-class examples well above chance: 0.72 on CUB-200 and 0.94 on CIFAR-100, versus chance levels of 0.06 and 0.10. This provides direct numerical evidence that the protected geometry captures task-relevant structure, addressing the concern that the anchored directions might be arbitrary or unrelated to prior-task knowledge. We also provided complementary image-based visualizations in the paper.

  • Empirical measurements of theoretical measures and comparisons with other geometries: We showed a very interesting aspect – SAGA’s benefit is not simply lower classifier leakage. Despite leakage comparable to PCA and higher than classifier-SVD, SAGA preserves local SAE support structure best and consistently reduces drift outside the protected subspace. It also maintains CKA comparable to PCA and achieves the lowest forgetting on CUB-200, with competitive results on CIFAR-100-224, showing unique advantage of SAGA for protecting both the local and global structures in representations at the same time. These strengthen our theoretical analyses of representation change.

  • Resource cost: We expanded storage and runtime accounting, making the method’s overhead directly accessible and addressing the concern that its practical cost was unclear.

We will be more than happy to engage in the discussion while the rebuttal is still open and to further clarify these aspects.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.