acceptodds
Under review as a conference paper at ICLR 2027

Masking Spurious Concepts via Sparse Autoencoders for Efficient Bias Mitigation

Abstract

Deep neural networks trained via Empirical Risk Minimisation (ERM) are prone to relying on spurious correlations: confounding features such as backgrounds that are statistically associated with a class label during training but are not causally relevant. When these correlations do not hold at test time, performance degrades on subgroups without spurious features. Existing mitigation methods either (a) require model retraining or large group label annotations, or (b) modify only the final classifier head, leaving spurious information encoded in the latent space and thus exhibiting a trade-off between worst-group and average accuracy. To circumvent these limitations, we propose SPOCK, a bias mitigation method that uses a Sparse Autoencoder to decompose latent representations into sparse concepts and learns a mask that suppresses spurious concepts while preserving those useful for the downstream task. By filtering concepts before reconstructing the representations, SPOCK reduces the model's reliance on spurious correlations while keeping the existing classifier fixed. In this way, SPOCK is compute- and label-efficient, requiring only a handful of group-annotated validation examples and avoiding end-to-end retraining. Our results across four synthetic and real-world datasets show that SPOCK substantially improves worst-group accuracy over ERM while maintaining high average accuracy, offering a more favourable trade-off than competing bias mitigation approaches.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
spurious correlations, sparse autoencoders, model robustness, bias mitigation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_SaaWOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies spurious correlations in pretrained models. The key idea is: instead of retraining the model or changing the classifier, identify and suppress spurious concepts in the latent representation before classification.

SPOCK differs from last-layer retraining methods because those methods leave the spurious information encoded in the latent representation and only adapt the final classifier. In contrast, SPOCK directly modifies the latent representation while keeping the original classifier fixed.

SPOCK assumes that the latent representation (z) is dense and entangled. The authors therefore train a Sparse Autoencoder (SAE) on the model's latent representations, which maps them into a much larger set of sparse concepts.

After training the SAE, SPOCK learns a mask over the SAE concepts. The mask suppresses some concepts and preserves others. The masked concept representation is then decoded back into the original latent space and passed to the original classifier.

The mask is learned using a small manually constructed group-balanced validation set, where examples from different groups are balanced so that the model does not simply preserve concepts associated with the majority groups.

Overall, evaluation shows that SPOCK improved worst-group accuracy (WGA) while maintaining relatively high average accuracy.

Critical strengths
  • No last-layer retraining required. A key strength of SPOCK is that it can mitigate spurious correlations without retraining or modifying the original classifier head. Instead, it learns a mask over the SAE concepts and applies it before the fixed classifier, making the approach lightweight and potentially more efficient than methods that require last-layer retraining. This is consequential because it provides a lightweight post-hoc approach that improves worst-group accuracy without requiring expensive end-to-end retraining.
  • Broad experimental evaluation. The paper evaluates SPOCK across multiple datasets and model architectures, including Waterbirds, CelebA, BAR, and MultiNLI, and compares against several strong baselines. The experiments also examine label efficiency, worst-group versus average accuracy, and whether the learned mask actually suppresses spurious information.
  • Label efficiency. The experiments show that SPOCK can achieve meaningful improvements in worst-group accuracy with only a very small group-annotated validation set, while maintaining relatively high average accuracy. This is practically important because obtaining group annotations for the full training set can be expensive, and the results across four datasets provide evidence that the approach is not limited to a single benchmark.
Critical weaknesses
  • The evaluation does not sufficiently establish the generality of the approach. The experiments use relatively small base models (ResNet-18 for vision and BERT for MultiNLI), while the proposed mechanism is motivated as a general post-hoc representation-level method. It would be important to evaluate whether the approach remains effective for larger and more modern pretrained models and whether its claimed advantage over last-layer retraining persists in those settings.
  • Dependence on predefined group annotations. SPOCK requires group-annotated validation examples to learn its concept mask. Consequently, its ability to mitigate spurious correlations depends on the groups represented in the validation set. Biases arising from unknown or unannotated attributes may remain unmitigated, since SPOCK is not explicitly designed to discover arbitrary sources of bias.
Questions for authors
  • How much does performance depend on SAE architecture, latent layer, expansion factor, sparsity coefficient, and mask regularization?
  • Would the approach work with much larger/modern foundation models rather than ResNet-18 and BERT?
  • How much additional inference cost does SAE encode -> mask -> decode introduce?
4 comments

Dear Reviewer SaaW,

Thank you for taking the time to review our work. We greatly appreciate you describing our experimental evaluation as "broad" and acknowledging its coverage across multiple benchmarks. Below, we address your questions and concerns.

(W1 / Q2) Does SPOCK remain effective on larger and more modern pretrained models, and does its advantage over last-layer retraining persist in these settings?

We thank you and reviewer QWw6 for raising questions about SPOCK's effectiveness on modern backbone architectures. Following your suggestions, we have added Appendix E, which details the results of evaluating vision and text benchmarks using CLIP ViT-B/32 and ModernBERT-base, respectively.

Our experiments show that SPOCK outperforms DFR in both WGA and average accuracy across the vision benchmarks and remains competitive with DFR on MultiNLI (Figure 10). Together with our original ResNet-18 and BERT experiments, these results demonstrate that SPOCK generalises across different vision and language backbones while maintaining a favourable WGA–average accuracy trade-off. Please see our response to reviewer QWw6 (Q2) for a detailed comparison with DFR across these benchmarks.

(W2) SPOCK depends on predefined group annotations and may not address unknown or unannotated biases.

We acknowledge SPOCK's reliance on group annotations to construct a balanced held-out set for mask optimisation and explicitly identify this as one of the main limitations in our paper. We further agree that SPOCK is not designed to discover unknown biases. However, we would like to clarify that SPOCK does not assume access to privileged group information beyond what existing methods already use, such as group-annotated validation data for hyperparameter selection.

Nevertheless, our results on BAR show that SPOCK performs well even when the group-balanced set is constructed using only limited knowledge of the distribution shift, without predefined group annotations. This suggests that SPOCK can remain effective even when group annotations are approximate, although identifying unknown biases remains an open challenge.

(Q1) How much does performance depend on SAE architecture, latent layer, expansion factor, sparsity coefficient, and mask regularization?

We previously investigated the effects of the SAE expansion factor and mask regularisation coefficient λm\lambda_m. Following your and reviewer jJ6k's suggestions, we additionally conducted an ablation on SAE sparsity coefficient λs\lambda_s. All results are reported in Appendix G, and we summarise the key findings below:

  • Mask regularisation (λm\lambda_m): Performance remains relatively stable across the vision benchmarks, whereas MultiNLI exhibits greater sensitivity (Figure 12).
  • SAE expansion factor: 8× and 12× exhibit the least compromise between WGA and average accuracy across benchmarks (Figure 13).
  • SAE sparsity (λs\lambda_s): Stronger regularisation does not consistently improve performance (Figure 14).

We do not ablate the choice of layer in the backbone architecture, as our method builds on the premise that last-layer representations encode both task-relevant and spurious information [3], as exploited by DFR, AFR, and other LLR-based methods. We also focus on the last layer to ensure a fair comparison with these methods and demonstrate the effectiveness of bias mitigation without retraining the classification head.

For the SAE architecture, we adopt a standard single-hidden-layer design commonly used in the mechanistic interpretability literature for concept discovery [1,2]. The primary purpose of using an SAE is to obtain an overcomplete representation that facilitates the separation of task-relevant and spurious concepts, which is crucial for mask learning (see Figure 8 for experiments showing this). As acknowledged in our limitations, SAEs are not guaranteed to discover semantically disentangled concepts, and our findings are supported by empirical evidence rather than a theoretical guarantee.

(Q3) How much additional inference cost does SAE encode -> mask -> decode introduce?

We thank you and reviewer jJ6k for raising questions about SPOCK's inference overhead. Following your suggestions, we have added Appendix F, which reports inference latency (ms) and peak GPU memory (MiB).

Our results show that the relative inference latency and peak gpu memory overhead decreases with increasing batch size (Figure 11). For instance, for ResNet-18, the latency overhead decreases from 4.9% to 1.4%, while the memory overhead decreases from 16.2% to 1.7% as batch size increases from 1 to 128.

Overall, these results show that SPOCK introduces minimal inference latency overhead, particularly at larger batch sizes. Please see our response to reviewer jJ6k (Q4 / W4) and Appendix F, where we report a detailed comparison of inference latency and peak GPU memory across different batch sizes.

Thanks for your response and additional experiments. These address most of my concerns, particularly regarding newer backbones and inference overhead.

My main remaining concern is the reliance on group information to learn the mask. The BAR results are promising, but it remains unclear how well SPOCK handles unknown or unannotated spurious attributes. But I find the approach interesting, and the additional experiments have strengthened the paper.

Thank you very much for your follow-up and for taking the time to engage in the rebuttal process. We are glad that our responses and additional experiments have addressed several of your concerns.

We appreciate your acknowledgement of SPOCK's performance on BAR, even with approximate group balancing. As previously indicated in our rebuttal, one limitation of our method is its reliance on group annotations to construct a balanced held-out set for mask optimisation, which we also acknowledge in the paper.

Nevertheless, our annotation-budget experiment (Figure 5) demonstrates that SPOCK can achieve competitive performance with only 1 example per group on CelebA, reducing its reliance on extensive group-annotated data. We agree that extending SPOCK to settings with unknown spurious attributes is a promising direction for future research, particularly by investigating whether SPOCK's learned masks can help identify additional sources of spurious correlation and subsequently guide more informed data collection and augmentation.


If you have any further questions or concerns, please do not hesitate to let us know. Thank you once again for engaging in this discussion and for your valuable feedback.

Reviewer_QWw6OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
3
Read review
Paper summary

This paper targets mitigating spurious correlations and improving the model's performance on subgroups without spurious features. It summarizes that existing methods either require model retraining / large group label annotations, or only modify the classification head and leave the latent feature space intact. To address this issue, this paper introduces SPOCK, which utilizes a sparse autoencoder and learns a task to suppress spurious concepts. Experiments are conducted on four datasets to verify the effectiveness.

Critical strengths

The idea that distinguishes spurious concepts from task-relevant concepts using sparse autoencoders and learning masks to guide the reconstructed representation is interesting. This could be an effective way to mitigate the spurious correlation issue from the perspective of latent representations instead of classification heads, and it requires only a small group-annotated held-out set.

Critical weaknesses
  • This paper summarizes the shortcomings of existing works as (1) requiring model retraining or large-group annotations; (2) only modifying the classifier head instead of feature representations. However, it does not analyze how such weaknesses would affect the model's ability to mitigate spurious correlations. Therefore, the motivation is somewhat limited.
  • What's the rationale for Sparse Autoencoder Training and Mask learning in mitigating the spurious concepts? How to guarntee that they would successfully distinguish spurious concepts from core task-relevant concepts? It is not quite clear.
Questions for authors
  • What is the approximate wall-clock time and computational cost of SPOCK?
  • How about the effectiveness of SPOCK on other models? It is encouraged to include more advanced backbones to demonstrate the efficacy of the proposed method, such as other CNN architectures, ViT-based backbones, and the latest text models.
4 comments

Dear Reviewer QWw6,

Thank you for your valuable feedback. We are glad you found our paper easy to follow and recognised key strengths of SPOCK, including its "efficiency with limited group annotations" and its ability to "avoid backbone retraining". Below, we address your questions and concerns.

(W1) Why is SPOCK needed when existing methods already mitigate spurious correlations?

We appreciate the importance of correctly positioning our work relative to existing methods. As already discussed in the Introduction and Related Work sections, we motivate SPOCK by identifying the limitations of three existing classes of bias mitigation methods. We summarise these below and explain how SPOCK addresses them:

  1. Group-Supervised Methods (e.g., Group DRO): require group annotations for the entire training set and explicitly use them to improve performance on under-represented groups.

    • Limitation: obtaining group annotations at scale is costly and prone to annotation errors; optimising for worst-group performance can come at the expense of average accuracy, introducing a trade-off between the two.
    • SPOCK: requires only a small group-annotated held-out set; remains effective even with as few as 2 examples per group on Waterbirds, while maintaining competitive average accuracy (Figure 5).
  2. Group-Unsupervised Methods (e.g., LfF, JTT): do not require group annotations during training; typically work by inferring pseudo-group labels followed by retraining.

    • Limitation: retraining or modifying the model is computationally expensive (Figure 4); these methods are highly sensitive to hyperparameter selection.
    • SPOCK: does not require retraining the backbone or classifier head; incurs lower training costs (Figure 4); uses a single regularisation hyperparameter, λm\lambda_m, to which performance remains relatively stable (Figure 12).
  3. Last-Layer Retraining Methods (e.g., DFR): retrain the classification head using a small held-out dataset that is either explicitly or implicitly group-balanced.

    • Limitation: leaves the original learned representations unchanged and modifies the classifier decision boundary to better fit under-represented groups, introducing a trade-off between worst-group and average accuracy (Figure 1).
    • SPOCK: directly modifies the learned representations by masking latent concepts (Figure 1); improves worst-group accuracy while maintaining high average accuracy (Figures 3 and 10).

(W2) What's the rationale for SAE training and mask learning in mitigating the spurious concepts?

SPOCK builds on the premise that, despite the presence of spurious correlations, DNNs capture both spurious and task-relevant concepts in their latent representations [2]. Our goal is to distinguish spurious concepts from task-relevant concepts and ultimately suppress the former to reduce reliance on spurious correlations during prediction.

However, due to superposition [1], individual latents can be polysemantic, i.e., they encode multiple unrelated concepts. As a result, suppressing individual latents to remove spurious concepts could also erase useful task-relevant information, thereby hurting performance. Therefore, we use an SAE, an unsupervised approach that aims to disentangle a model's latent representations into distinct concepts. We hypothesise that SAEs provide a representation in which spurious and task-relevant information can be more selectively suppressed or preserved. We can see this in Figure 8, where we show that masking SAE concepts achieves higher WGA compared with directly masking the original dense latent representations.

Mask learning is then necessary to identify which concepts should be suppressed and which preserved. We therefore learn a sparse mask over the SAE concepts using a small group-balanced held-out set, to retain concepts that generalise well across groups and suppress the spurious concepts.

How to guarantee that they would successfully distinguish spurious concepts from core task-relevant concepts?

There is no guarantee that an SAE disentangles dense latent representations into distinct task-relevant and spurious concepts. However, to test our hypothesis that concepts suppressed by SPOCK capture more spurious information than those preserved, we train separate linear probes on the most suppressed and preserved concepts. As shown in Figure 6, probes trained on the suppressed concepts achieve higher accuracy in predicting spurious attributes than those trained on preserved concepts. This suggests that the suppressed concepts encode more information about spurious attributes, providing empirical support for our hypothesis. However, this does not imply that suppressed concepts contain no task-relevant information. Since spurious attributes are correlated with task labels in the training distribution, concepts encoding spurious information may also contain information predictive of task labels, and vice versa.

(Q1) What is the approximate wall-clock time and computational cost of SPOCK?

We thank you and the other reviewers for raising questions about SPOCK's runtime costs and efficiency. Following your suggestions, we have added Appendix F, which details the computational costs during training and inference.

Below, we summarise the total training time across datasets, which includes training an ERM checkpoint, followed by SAE training and mask optimisation ( TERM+TSAE+TMASKT_{ERM} + T_{SAE} + T_{MASK}):

DatasetERM (s)SPOCK (s)
Waterbirds28.08 ± 7.7535.26 ± 0.07
CelebA213.41 ± 42.67229.52 ± 0.03
BAR14.28 ± 3.0618.09 ± 0.06
MultiNLI1363.92 ± 1.041518.60 ± 0.05

Overall, SPOCK incurs modest additional training costs over ERM (up to 1.3×), while avoiding the expensive backbone retraining required by other group-unsupervised methods, which can incur up to 6× the ERM runtime (Figure 4). For a detailed breakdown of additional training time starting from a fixed ERM checkpoint, inference latency, and peak GPU memory, please see our response to reviewer jJ6k (Q4 / W4) and Appendix F.

(Q2) Does SPOCK remain effective with more advanced CNNs, ViT-based backbones, and the latest text models?

Following your suggestion, we further evaluated SPOCK using CLIP ViT-B/32 for vision benchmarks and ModernBERT-base for MultiNLI, and reported our findings in Appendix E. Our new results suggest that SPOCK consistently outperforms DFR, a state-of-the-art last-layer retraining method, on all three vision benchmarks in terms of both worst-group and average accuracy, while remaining competitive on MultiNLI. Below, we summarise the results across benchmarks:

MethodWaterbirds Worst (%)Waterbirds Average (%)CelebA Worst (%)CelebA Average (%)BAR Worst (%)BAR Average (%)MultiNLI Worst (%)MultiNLI Average (%)
ERM (baseline)31.67 ± 3.6384.64 ± 0.6232.96 ± 4.3295.13 ± 0.0352.30 ± 5.3388.58 ± 0.8666.67 ± 4.8486.43 ± 0.22
DFR81.08 ± 2.4282.81 ± 1.5480.19 ± 1.8882.14 ± 2.1580.78 ± 2.1694.12 ± 0.4375.58 ± 3.5885.06 ± 0.49
SPOCK83.76 ± 0.6387.68 ± 0.7486.55 ± 2.6490.52 ± 0.8885.23 ± 0.8495.21 ± 0.2873.95 ± 1.3285.58 ± 0.20

These results demonstrate that SPOCK generalises well across different vision and language backbones, achieving a favourable trade-off between worst-group and average accuracy.

References

  1. Huben, Robert, et al. "Sparse autoencoders find highly interpretable features in language models." ICLR, 2024.
  2. Kirichenko, Polina, Pavel Izmailov, and Andrew Gordon Wilson. "Last layer re-training is sufficient for robustness to spurious correlations." ICLR, 2023.

Thanks for the detailed rebuttal, which helps address my concerns. I have increased the score to 3 and support the acceptance of this paper. Please ensure that the provided discussion on distinctions between existing methods, rationale, and new results are included in the updated version of the paper.

Thank you very much for engaging with our rebuttal and reconsidering your assessment. We appreciate your support for the acceptance of our paper. Should the paper be accepted, we will ensure that all relevant discussions, clarifications and additional experimental results from the rebuttal are incorporated into the camera-ready version.

Reviewer_jJ6kOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SPOCK to mitigate spurious correlations using a sparse autoencoder and a learned feature mask. The backbone and classifier head are kept frozen. The authors first train the autoencoder on training representations, then use a group-balanced held-out set to learn which features to suppress. The masked features are decoded and fed to the original classifier. Experiments on Waterbirds, CelebA, BAR, and MultiNLI compare SPOCK with ERM, robustness methods, and last-layer retraining. SPOCK improves worst-group accuracy over ERM. Compared with DFR, it improves both average and worst-group accuracy on CelebA and MultiNLI, while trading some worst-group or worst-class accuracy for higher average accuracy on Waterbirds and BAR. The paper also studies group-label budgets, direct feature masking, and representation probes.

Critical strengths
  1. SPOCK improves both average and worst-group accuracy over DFR on CelebA and MultiNLI. On CelebA, the two accuracies increase from 85.16% to 89.92% and from 79.79% to 84.65%, respectively. The methods use the same held-out set configuration and model selection procedure.

  2. The improvements on CelebA and MultiNLI remain with fewer group annotations in Figure 5. The method also avoids backbone retraining and takes less training time than JTT and LfF in Figure 4.

Critical weaknesses
  1. Figure 8 compares SAE masking with masking the original features. However, the SAE also learns a nonlinear feature map and expands the number of coordinates eightfold. This comparison cannot tell whether sparsity is needed or whether the new feature map is enough. A dense autoencoder with the same width, reconstruction training, and mask objective is missing. Keeping the classifier head frozen does not resolve this, since the features given to the head have changed.
  2. On BAR, average accuracy improves from 78.80% with DFR to 85.39%, but worst-class accuracy drops from 62.46% to 60.00%. Waterbirds has the same tradeoff. I would be careful about claiming a better accuracy tradeoff from these selected configurations. Both methods are selected by validation worst-group accuracy, and Figure 9 only sweeps SPOCK's regularization. We do not know whether DFR can reach similar average accuracy if its operating point is changed.
  3. The probe results on MultiNLI do not show much separation between suppressed and preserved concepts, and direct masking performs comparably to SAE masking. The authors acknowledge the probe result, but it leaves the role of the SAE unclear in this setting. On the vision datasets, the separation is clearer, though only spurious-attribute information is measured. Without also probing task information, it is hard to tell how selective the removal is.
  4. Figure 4 includes the common ERM training cost. For an existing model, I am more interested in the additional cost relative to DFR, which also keeps the backbone frozen and does not add an autoencoder. The adaptation time, inference latency, and memory should be compared separately.
Questions for authors
  1. How does SPOCK compare with a dense autoencoder using the same width, reconstruction training, and mask objective?

  2. Does SPOCK still have higher average accuracy than DFR at matched worst-group accuracy and annotation budget?

  3. How much task information remains in the preserved and suppressed concepts? Can the probes separate task information from spurious-attribute information?

  4. Starting from the same checkpoint, how much additional training time, inference latency, and memory does SPOCK require compared with DFR?

8 comments

Dear Reviewer jJ6k,

Thank you for your valuable feedback. We are glad you found our paper easy to follow and recognised key strengths of SPOCK, including its "efficiency with limited group annotations" and its ability to "avoid backbone retraining". Below, we address your questions and concerns.

(Q1 / W1) How does SPOCK compare with a dense autoencoder using the same width, reconstruction training, and mask objective?

In our original experiments, we used a fixed SAE sparsity penalty λs\lambda_s for vision and language backbones, based on initial exploration. Our motivation for using an SAE was to decompose dense, entangled representations into disentangled concepts that could be preserved or suppressed during mask learning. Since SPOCK performed well under these settings, we did not investigate the sensitivity to λs\lambda_s in our initial evaluation.

We have now added an ablation in Appendix G that varies λs∈[0.0,1.0]\lambda_s \in [0.0, 1.0] while keeping the remaining architecture, training configuration, and mask-learning objective fixed. Following your suggestion, we also include λs=0\lambda_s = 0 to evaluate an autoencoder without explicit sparsity regularisation. Our results show that weaker sparsity regularisation, including λs=0\lambda_s = 0, generally leads to better performance, whereas stronger sparsity penalties do not consistently improve performance. Nevertheless, sparsity may still benefit interpretability, allowing future work to analyse SPOCK's learned masks and potentially guide more informed data collection or augmentation.

Overall, these findings suggest that explicit sparsity regularisation is not essential for SPOCK's observed robustness improvements. We agree that Figure 8 does not necessarily establish the benefits of sparsity itself. We have therefore revised its discussion to emphasise that masking overcomplete SAE representations results in higher WGA than directly masking the original dense latent representations.

(Q2 / W2) Can DFR achieve a similar WGA–average accuracy trade-off to SPOCK at a different operating point?

We would like to clarify that DFR uses a regularisation hyperparameter CC, which controls the inverse regularisation strength. As with SPOCK and the other baselines, we select the final configuration based on the highest validation WGA. Both SPOCK and DFR use the same fixed ERM checkpoint and held-out data splits, including the same random seeds, ensuring a consistent group-annotation budget across methods. As detailed in Appendix C, we tune the methods following the guidelines of their respective original papers.

While WGA cannot be controlled directly, we can compare methods in settings where they achieve similar observed WGA. To investigate this, we revisited the results underlying the group-label efficiency plots in Figure 5 and identified configurations where SPOCK and DFR achieve approximately matched WGA at the same annotation budget. We then compared their corresponding average accuracies.

DatasetAnnotation BudgetDFR WGA (%)SPOCK WGA (%)DFR Avg. Acc. (%)SPOCK Avg. Acc. (%)Δ Avg. Acc.
Waterbirds10%76.05 ± 5.8575.51 ± 4.1182.35 ± 2.5388.54 ± 2.82+6.19
Waterbirds20%78.64 ± 4.9479.06 ± 3.1983.71 ± 2.5388.13 ± 1.25+4.42
BAR20%44.62 ± 11.3644.62 ± 7.4668.01 ± 4.9769.96 ± 2.78+1.95
BAR30%52.96 ± 7.3552.58 ± 6.5371.87 ± 2.8071.50 ± 2.24−0.37

As shown above, at approximately matched mean WGA and annotation budgets, SPOCK consistently achieves higher average accuracy than DFR on Waterbirds. On BAR, the differences are smaller and mixed. As BAR lacks explicit group annotations, we use class labels as a proxy to construct an approximately balanced held-out set, under the assumption of a distribution shift between training and evaluation. This could explain why SPOCK shows smaller and less consistent gains over DFR on BAR compared with Waterbirds.

We also explored alternative mask sparsity coefficients λm\lambda_m for SPOCK on CelebA and MultiNLI, beyond those selected by validation WGA, to identify configurations with WGA comparable to DFR. However, across all configurations considered, SPOCK maintained higher WGA than DFR's best observed WGA, preventing us from making the requested matched-WGA comparison.

Finally, we thank the reviewer for highlighting the reported average accuracy improvement on BAR. Upon revisiting our results, we identified that SPOCK's average accuracy was mistakenly reported as weighted accuracy (85.39%), resulting in a zero SD. We sincerely apologise for this oversight and have corrected Table 2 and the corresponding trade-off plot in Figure 3. Nevertheless, we emphasise that this update does not change any of the claims made in our paper or general trends observed across settings. We have also meticulously reverified all our results and ensured that there are no other discrepancies elsewhere in the paper.

(Q3 / W3) How much task information remains in the preserved and suppressed concepts?

We thank the reviewer for the suggestion. In Section 5.5, we train linear probes on a group-balanced validation subset to predict spurious attributes. We extend this analysis to predict task labels using the same sets of preserved and suppressed concepts to investigate whether preserved concepts capture more task-relevant information than suppressed concepts. These additional results are reported in Figure 9 in Appendix D.

Our results show that the most preserved concepts (top-kk) generally achieve higher task-probe accuracy than the most suppressed concepts (bottom-kk). This distinction is more pronounced on Waterbirds and CelebA, consistent with the spurious-attribute probes in Figure 6, which show the opposite trend. This provides further evidence that SPOCK preferentially preserves task-relevant information while suppressing spurious information on the vision benchmarks.

On MultiNLI, the difference between the two sets of concepts is less pronounced, consistent with our original spurious-attribute probing results. We leave a better understanding of this behaviour to future work.

Finally, we acknowledge that these findings are based on empirical evidence and do not guarantee that SAEs disentangle dense latent representations into distinct task-relevant and spurious concepts, as clearly indicated in our limitations in Section 6.

(Q4 / W4) How much additional training and inference overhead does SPOCK incur over DFR?

Following your suggestion, we have added Appendix F, which details the computational costs incurred during training and inference. To ensure consistent comparisons, we also re-ran the training-time measurements on the same hardware used for inference and updated Figure 4 accordingly.

Training cost. Starting from the same fixed ERM checkpoint, SPOCK trains an SAE on the training representations and subsequently optimises a mask over SAE concepts, whereas DFR only retrains the classification head. We summarise the additional wall-clock training time required by both methods across datasets below:

DatasetDFR (s)SPOCK (s)
Waterbirds2.27 ± 0.027.19 ± 0.07
CelebA7.58 ± 0.0316.12 ± 0.03
MultiNLI122.77 ± 0.81154.68 ± 0.05
BAR1.04 ± 0.023.81 ± 0.06

SPOCK requires up to 3.7× the additional training time of DFR on BAR, the smallest dataset, but only 1.3× on MultiNLI, the largest. Nevertheless, its total training time (Figure 4) remains within 1.3× that of ERM, compared with 1.1× for DFR, representing a modest additional cost for providing a favourable trade-off between worst-group and average accuracy.

Inference overhead. We additionally measure inference latency and peak GPU memory across different batch sizes. Detailed results are presented in Figure 11 in Appendix F, and we summarise them below:

BackboneBatch SizeLatency (ms) BaselineLatency (ms) SPOCKPeak GPU Memory (MiB) BaselinePeak GPU Memory (MiB) SPOCK
ResNet-1811.331 ± 0.0011.396 ± 0.00098.87114.90
ResNet-1812827.227 ± 0.06627.609 ± 0.014959.32975.35
BERT13.400 ± 0.0023.548 ± 0.003436.98473.91
BERT128189.690 ± 0.184190.066 ± 0.0161340.291377.22

As also evident in Figure 11, the relative inference latency overhead decreases with increasing batch size. For ResNet-18, it decreases from 4.9% at batch size 1 to 1.4% at batch size 128, while for BERT, it decreases from 4.4% to 0.2%, respectively. Similarly, the peak GPU memory overhead decreases from 16.2% to 1.7% for ResNet-18 and from 8.5% to 2.8% for BERT.

Overall, these results show that although SPOCK requires additional training time compared with DFR, it introduces minimal inference latency and memory overhead, particularly at larger batch sizes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.