acceptodds
Under review as a conference paper at ICLR 2027

Pseudo group-guided and reliability-aware debiasing framework for unsupervised bias mitigation

Abstract

Deep neural networks often exploit spurious correlations between target labels and task-irrelevant attributes, leading to substantial performance degradation on samples that deviate from dominant spurious patterns. To address this issue, we propose a selective and reliability-aware debiasing framework for settings where spurious attributes are unavailable. Our approach extends representation-based pseudo-group discovery through Gaussian Noise Intervention (GNI), which evaluates the stability of intra-class geometric evidence under bounded local perturbations and derives a conflict score for selective pseudo-group inference. Within each target class, samples are divided into pseudo-aligned, pseudo-conflicting, and uncertain groups. Retained pseudo-conflicting samples are further assigned reliability scores based on multi-view consistency and class-wise ranking. These pseudo-group assignments and reliability estimates are incorporated into a reliability-weighted multi-level debiasing objective, where they play distinct roles in emphasizing bias-conflicting samples and modulating their debiasing contributions. To further mitigate residual shortcut influence, we employ trained bias experts to perform Debiased Correction Of Residual Effects (D-CORE) at inference stage. Experiments on medical and natural-image benchmarks demonstrate improved worst-group robustness across most evaluated settings while maintaining competitive average accuracy.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Debiased Classification, Dataset Bias, Gaussian Noise Intervention, Pseudo Group

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

I read and evaluated the manuscript all by myself, the review is my own analysis and academic judgment. I only use chatgpt for language polishing to improve the readability.

My original text as as follows:

This paper proposes an unsupervised debiasing method. It combines Gaussian Noise Intervention (GNI) for pseudo-group discovery, reliability-aware multi-level debiasing, and D-CORE for inference-time residual-shortcut correction. Experiments on four benchmark datasets shows improved worst-group accuracy, it doesn’t need spurious-attribute annotations during training and model selection. However, the framework is designed very complex, the D-CORE is limited to binary classification, which limits its practical application.

  1. Debiasing without access to group annotations is a practical and interesting topic.

  2. Using trained bias experts to correct residual shortcut effects at inference time may be a good and novel direction for adaptive debiasing.

1.The motivation of the paper, the necessity of the designed techniques are insufficiently justified. The paper does not discuss the limitations of existing bias-mitigation methods, which the paper aims to address.

2.论文的关键模块在设计、理论支撑和实验验证方面不足。it is not sure whether the theoretical of GNI translate into more accurate pseudo-group prediction, whether D-CORE can isolate and correct reliance on residual shortcut.

3.The framework involves multiple components, loss terms, and hyperparameters, raising concerns about implementation complexity and tuning costs. Moreover, D-CORE is currently restricted to binary classification, limiting the method's broader applicability.

  1. 摘要关于本文的motivation过于笼统了. Many bias-mitigation methods already exist, the paper should clarify what kind of limitation existing methods have, what kind of limitation it try to address, what advantages it brings compare to existing methods.

  2. Figure 1, caption, "correctly classified" is used without specifying what a correctness refers to, does it means correct target-label prediction or does it means correct pseudo-group assignment. The evaluation criterion used should be clearly descripted.

  3. Related work is not up to date, more recent works should be discussed.

  4. Since the work builds on existing method NSF, it should be introduced in preliminary for easy understanding.

  5. 符号不统一,different parts of the paper use different notations (rho and l) for the layer index.

  6. The authors should clarify, the hyperparameter selection, evaluate the practical tuning costs.

  7. Concern about GNI technique: Reduce the variance of geometric statistic does not necessarily improve pseudo-group prediction accuracy because the statistic may remain systematically biased. There is a gap between the theoretical analysis and the practical implementation, please discuss it. The motivation that multi-view stability improves pseudo-group discovery need further verification.

  8. Can the method maintain normal performance on data without spurious correlations? a method that degrades performance in this setting is not practical.

  9. D-CORE 将bias expert对factual representation的响应与对zero-feature reference的响应之间的差异,理解为residual-shortcut signal。However,由于偏置专家是使用target labels而非伪相关属性annotation进行监督训练,他会同时捕获shortcut-related and target-related information。when spurious correlations is weak,target-related information甚至可能占据主导地位。而用zero vector replace input representation 不一定能够分离出伪相关特征所产生的贡献。 相关说法需要更充分的理论依据,建议设计对应的 controlled study区分residual-bias correction与传统auxiliary-model correction的关系与区别。

  10. 它在 multi-shortcut 情况下如何?, because it’s the case in real-world application.

Critical strengths
  1. Debiasing without access to group annotations is practically appealing.

  2. Using trained bias experts to correct residual shortcut effects at inference time is a potentially promising direction for adaptive debiasing.

Critical weaknesses

1.The motivation and necessity of the proposed design are insufficiently justified. The paper does not clearly establish the specific limitations of existing bias-mitigation methods that the proposed framework aims to address.

2.The theoretical justification and empirical validity of key components remain insufficiently established. In particular, it is unclear whether the theoretical properties of GNI translate into more accurate pseudo-group discovery, and whether D-CORE genuinely isolates and corrects residual shortcut effects.

3.The framework involves multiple components, loss terms, and hyperparameters, raising concerns about implementation complexity and tuning costs. Moreover, D-CORE is currently restricted to binary classification, limiting the method's broader applicability.

Questions for authors
  1. The motivation stated in the abstract is too generic. Many bias-mitigation methods already exist; the paper should clarify which specific limitation of existing methods it addresses and what advantage the proposed method offers over them.

  2. The caption of Figure 1 uses "correctly classified" without specifying what correctness refers to—whether correct target-label prediction or correct pseudo-group assignment. The authors should explicitly clarify the evaluation criterion.

  3. Related work is insufficient; more recent works should be discussed.

  4. Since the work builds on existing NSF, NSF should be introduced as a preliminary so that the paper is easier to follow.

  5. The layer index is denoted by both rho and l in different parts of the paper; this notation should be unified.

  6. The authors should clarify the hyperparameter selection process and evaluate the associated tuning costs in practice.

  7. Concern about GNI: Reducing the variance of the geometric statistic does not necessarily improve pseudo-group assignment accuracy, since the statistic may remain systematically biased. The authors should clarify the gap between the theoretical analysis and the practical implementation, and verify whether multi-view stability improves pseudo-group discovery.

  8. Can the method maintain normal performance on data without spurious correlations? This matters because a method that degrades performance in this setting would have limited practical value.

  9. D-CORE interprets the difference between a bias expert's responses to the factual representation and to a zero-feature reference as a residual-shortcut signal. However, since the bias experts are trained with target labels but without explicit spurious-attribute supervision, they may capture both shortcut-related and target-relevant information, with the latter potentially dominating when spurious correlations are weak. Moreover, replacing the input representation with a zero vector does not necessarily isolate the contribution of spurious features. The authors should provide stronger justification for this interpretation and conduct controlled experiments to distinguish genuine residual-bias correction from generic auxiliary-model correction.

  10. How robust is the method under multiple-shortcut settings, which are more common in real-world applications?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 3 on Method Motivation and Validation

W1&Q1&Q2&Q4&Q5: We thank the reviewer for the feedback on the manuscript’s structure and presentation. We revised the abstract and replaced “correctly classified” with “correct target-label predictions” in the Figure 1 caption. We expanded the related work to compare D-CORE with recent counterfactual debiasing methods, introduced NSF in the preliminaries, and standardized the notation for layer indices throughout the manuscript. These revisions strengthen the motivation for our framework and clarify which limitations of existing methods it addresses. W2&Q7&Q9: We thank the reviewer for the questions. NSF’s sign-based grouping can fail under weak bias. To address this limitation, GNI uses intra-class ranking, which can correctly separate two groups under weaker bias conditions in our theoretical model. Using multiple perturbed views can further improve pseudo-group quality. Under our assumptions, the upper bound on the probability of reversing score order decreases exponentially with the number of independent views. Detailed proofs are provided in Appendix A. We added clean-score baseline that applies the same intra-class ranking to noise-free scores with all other settings held fixed (Appendix D, Table 4). We also varied the number of perturbed views to assess effects on pseudo-group quality and downstream performance (Appendix D.1, Table 5).In Appendix B.1, we added analysis of residual correction under controlled interventions on spurious features. These results show that zero-feature responses provide an effective reference for isolating residual shortcut signals and that aggregating these signals to correct predictions improves prediction consistency under the interventions. To compare D-CORE with a generic auxiliary-model correction in our seven image-only classification settings, we used the official CF-VQA code to adapt its correction and substitute it for D-CORE, keeping training hyperparameters fixed. D-CORE achieves higher WGA and mean accuracy in all seven settings (Appendix D.4, Table 8). W3&Q6: We thank the reviewer for the questions. We also added hyperparameter sensitivity analysis (Appendix D1&D7, Table5&Table12). Performance is sensitive to the voting threshold, which determines which samples are assigned to pseudo-groups and thus affects their quality. In Appendix I, we compare total runtime and peak GPU memory usage with those of baselines. On MIMIC-CXR+NIH at bias ratio 0.90, our method takes longer to run but has the lowest peak GPU memory usage among the competing methods. It also achieves the best WGA among all competing methods. We will focus on reducing training time in future work. In Appendix B.2, we prove that, under the stated assumptions, D-CORE can also reduce the model’s reliance on spurious features in multiclass classification. Q8&Q10:We thank the reviewer for raising this question about practical applicability. We evaluated our method on an unbiased variant of Waterbirds (Appendix H, Table 16). The results show that our method maintains strong predictive performance in the absence of spurious correlations. We also evaluated our method on HAM10000 with age and sex as spurious attributes (Appendix G, Table 15). Our method improves both worst-group and overall accuracy, demonstrating its ability to handle multiple coexisting shortcuts in this real-world image classification setting.

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

This paper studies how to reduce bias when no bias labels are available, not even for model selection. The method first splits each class into likely bias-conflicting, bias-aligned and uncertain samples. To do this it adds Gaussian noise to frozen ERM features several times and averages a distance-based score over the noisy copies. It then trains a network with bias and debiasing experts at two layers and a learned gate that mixes the main and debiasing outputs. Samples in the conflicting group get extra weight in the prediction loss, and in the debiasing loss this extra weight is scaled by a reliability score built from the share of their noisy copies that look conflicting and from their rank in that group. The bias experts are trained without group weights, and several extra losses are added, including a prototype loss. At test time, a step the authors call D-CORE compares each bias expert's margin on the real features with its margin on all-zero features, and subtracts a scaled, layer-weighted sum of these differences from the model's decision margin.

The evaluation covers seven binary settings. These are three bias ratios on MIMIC-CXR+NIH, two GbP bias directions, CelebA and Waterbirds. The method has the best worst-group accuracy in all of them, on average 2.37 points above the strongest competitor, and the authors read the ablations as showing that the parts complement each other.

Critical strengths
  • The setting is strict. No bias labels are used for training or for model selection, while methods such as JTT and DFR need a validation set with group labels.

  • The medical benchmarks are useful. MIMIC-CXR+NIH has three bias ratios and GbP has two bias directions, so one can see how methods behave as the bias gets stronger or changes direction. ERM drops to 5.76 worst-group accuracy at ratio 0.95, and on these medical settings the method's worst-group accuracy is 16 to 21 points higher than the reproduced NSF.

  • Putting borderline samples into an uncertain group, instead of forcing every sample into one of two groups, makes sense. The paper is also honest about the limits of its theory and says that its analysis of the noisy copies does not imply an improvement over an oracle noise-free statistic.

Critical weaknesses
  • I am not sure the added noise in GNI gives the benefit the paper describes. With a fixed distance metric and fixed class centers, the noise only shifts each sample's score by a random term whose distribution, in these binary tasks, depends on the class but not on the sample. So in expectation, the averaged score, the consistency and the reliability weight are all monotone in the original noise-free score, and none of them can rank two samples of a class in the opposite order. If this is right, the noisy copies do not add information beyond that score in expectation, which is consistent with the paper's own caveat.

  • D-CORE seems closely related to earlier methods that are not discussed. If the layer weights are the same for every sample, the zero-feature reference only adds a constant, and the tuned decision threshold absorbs it. What remains is subtracting a scaled bias-expert margin from the model's decision margin. Product-of-experts debiasing, RUBi and counterfactual debiasing in VQA all use a bias-only branch to correct the main model, and the last one subtracts the bias branch's effect at test time as D-CORE does. A discussion of these would help position the method. The search range for the strength of the correction is also not given.

  • The ablations do not yet make clear where the gains come from. Without D-CORE the model reaches 58.72 on MIMIC 0.90, below Ada-ABC at 65.76, so the lead seems to depend on the test-time correction. Removing the prototype loss costs about 20 points, more than removing D-CORE or setting all reliability weights to one, even though the method section presents it as an extra regularizer. Using only the consistency term or only the rank term as the reliability weight gives 42.88 and 44.48, well below the 55.80 obtained when all reliability weights are one. Since in expectation both depend only on the same noise-free score, as argued above, these large differences would be worth explaining. The only single-view row in the ablation also uses a binary split with an unstated threshold, so no ablation turns off the noise while keeping everything else the same. A comparison of Euclidean and Mahalanobis distance is also missing, and all ablations are run only on MIMIC 0.90.

Questions for authors
  • Could you run the same pipeline without noise, using the original score directly? If the results are similar, it would help to clarify what GNI adds.

  • Did the baselines get the same checkpoint and threshold selection on validation data? Results for your method on all seven settings with a fixed threshold of 0.5 and without D-CORE would also help.

  • Do the D-CORE layer weights depend on the sample? If not, how is D-CORE different from simply subtracting the bias-expert margin, and what values of the correction strength do you search over?

1 comment
AuthorsOpenReview ↗

Justification of number of Gaussian-perturbed views , Clarification and numerical performance of D-CORE layer weights and Extended ablation studies on additional datasets and components

W1&Q1:We thank the reviewer and agree that, with fixed metrics and class centers, zero-mean Gaussian noise preserves the clean score in expectation, adding no ranking information at the expected-score level. Our framework combines Gaussian-perturbed views with intra-class ranking, which relaxes NSF’s strong-bias condition for sign-based separation and supports population-level group ordering under weaker bias conditions. The full proofs are provided in Appendix A. We have also added experiments evaluating the effect of view count on method performance and pseudo-group quality, along with a clean-score ablation using the original noise-free score, with other settings held fixed. The results are reported in Appendix D, Tables 5 and 4, respectively. The analysis in Appendix A and the experimental results jointly support improved ranking stability and pseudo-group quality as the number of independent views increases. Under the stated assumptions, the theoretical analysis establishes an exponentially decreasing upper bound on the probability of reversing the score ordering. W2&Q3:We thank the reviewer for highlighting the importance of layer weights and suggesting these related methods. Section 3.4 defines sample-adaptive aggregation weights for residual-shortcut estimation. When zero-feature reference margins differ across layers, these weights produce a sample-dependent reference contribution that cannot generally be absorbed into a global threshold. We have expanded Related Work to discuss Product-of-Experts debiasing, RUBi, and counterfactual debiasing in VQA, clarifying their connections to and differences from D-CORE. Counterfactual Debiasing targets VQA, whereas our seven tasks involve image classification without question inputs. To address the reviewer’s concern, we adapted CF-VQA using its official code to replace D-CORE while retaining identical training hyperparameters, and the resulting comparison in Table 8 of Appendix D.4 shows that D-CORE achieves higher WGA and mean accuracy across all seven settings. Appendix C.3 details the correction-strength search range and validation-based selection procedure, including a WCA constraint relative to the uncorrected model. The selected strength is applied uniformly to all test samples, independently of test outcomes. W3:We have extended the ablations to all seven dataset settings used in this study, with results in Table 2 of Section 4.3. GNI with reliability weighting already performs well under weaker spurious correlations. Under stronger correlations, D-CORE complements training-stage debiasing by aggregating multi-layer bias-expert margin differences between factual features and zero-feature references into a sample-specific correction of the final decision margin. Prototype regularization complements reliability weighting with a shared class-specific reference formed from normalized pseudo-aligned and pseudo-conflicting centroids. Controlling each group’s contribution reduces aligned-group dominance, while moving-average updates stabilize the centroids. Prototype alignment draws both groups toward this reference, encouraging intra-class consistency and reducing separation associated with spurious attributes. Table 9 in Appendix D.5 supports this role through a comparison with and without prototype regularization, holding all other settings fixed. We acknowledge that the original reliability ablation conflated component removal with changes in exponentiation. The revised consistency-only and rank-only configurations retain their respective exponents to assess complementary cross-view agreement and intra-class position. These results and the clean-score ablation are reported in Appendix D, Table 4. With their respective exponents retained, both the consistency-only and rank-only configurations outperform uniform weighting, where all reliability weights are set to one. The Euclidean–Mahalanobis comparison is provided in Appendix D.6, Table 10. The results favor our regularized Mahalanobis metric, supporting the benefit of accounting for unequal feature scales and correlations that Euclidean distance does not capture. Q2: We thank the reviewer for requesting clarification. We followed official checkpoint- and threshold-selection procedures, retaining GERNE’s validation WCA criterion. Our method uses WCA, which requires only target labels, to accommodate scarce bias annotations in medical imaging. We also evaluated WGA-based checkpoint selection to assess group-informed validation. Appendix D.7, Table11 shows that validation group annotations improve model selection and performance. We also added experiments comparing fixed and validation-selected thresholds and evaluating performance without D-CORE. Appendix D.3,Table7 shows competitive performance at a fixed threshold of 0.5 and further gains from WCA-based threshold selection. Section 4.3, Table 2 shows strong performance on several datasets without D-CORE.

Reviewer_Nbx4OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper tackles group-robust learning without spurious-attribute annotations by proposing a selective and reliability-aware debiasing framework. The key idea is to extend representation-based pseudo-group discovery with Gaussian Noise Intervention (GNI) to assess the local stability of intra-class geometric evidence, enabling selective assignment into pseudo-aligned, pseudo-conflicting, and uncertain groups with reliability scores for the retained conflicting samples. These signals drive a reliability-weighted, multi-level debiasing stage (with bias/debiasing experts and an invariant-prototype regularizer), followed by an inference-time correction module (D-CORE) that uses bias experts to adjust decision margins. Experiments on four benchmarks (medical and natural-image) show consistent improvements in worst-group accuracy, with generally competitive average accuracy.

Critical strengths
  • The extension of NSF-style geometric evidence to a multi-view, stability-oriented GNI probe is conceptually interesting and practically motivated; it reduces brittleness of single-snapshot geometry while enabling selective (A/C/U) assignment.
  • Evaluations span multiple spuriousness types (source, gender, background) across four benchmarks and three bias ratios on MIMIC-CXR+NIH, reporting mean and worst-group accuracy over five seeds.
  • Ablations isolate contributions of selective GNI, reliability-aware training, expert placement, and D-CORE; fine-grained analyses in the appendix further examine prototype regularization and pseudo-group quality.
  • The analytical appendices (idealized properties of multi-view stability and D-CORE variance analysis) are well-scoped and help interpret design choices without over-claiming guarantees.
  • The framework’s improvements in worst-group accuracy while maintaining competitive mean accuracy would be of practical interest to the community, and the inference-time correction adds a complementary tool.
Critical weaknesses

Major Weaknesses

  1. Appendix A.3 shows that Gaussian perturbations preserve the clean relative-distance statistic in expectation. It remains unclear why adding perturbation noise should improve pseudo-group discovery compared with directly using the deterministic clean statistic. An explicit clean-score baseline and sensitivity analysis over the number of perturbation views are needed to isolate the contribution of GNI.

  2. The detailed component ablations are restricted to MIMIC-CXR+NIH at a bias ratio of 0.90. Important hyperparameters governing group selection, perturbation, reliability weighting, and inference correction are not sufficiently documented. Moreover, some improvements over the strongest baselines are small relative to their reported variability, making additional statistical validation important.

  3. D-CORE uses a zero-feature reference passed into bias experts; while effective empirically, the choice of “zero feature” as a counterfactual baseline is heuristic and may not align with a principled feature ablation.

Minor

  1. The paper cites several environment/pseudo-group inference and representation-level debiasing papers but omits direct experimental comparisons with influential recent group-free post-hoc or data-centric methods (LFR, EVaLS, SAGE, Evidential Alignment), which appear to be particularly relevant and competitive in the same benchmarks.

  2. The individual consistency-only and rank-only reliability formulations perform worse than uniform weighting in Table 3. The source of this non-monotonic behavior is insufficiently investigated.

  3. The method sacrifices average accuracy on some benchmarks, particularly Waterbirds, compared with the strongest average-accuracy baseline. This trade-off deserves more discussion.

  4. The paper selects checkpoints and D-CORE parameters by worst-class accuracy, which is non-standard relative to some baselines; it is unclear if all compared methods are re-tuned under the same selection protocol to ensure a fair comparison.

  5. Computational overhead is not quantified (GNI multi-view probing, multi-branch experts, moving-average prototypes, and D-CORE inference); run-time or memory costs versus baselines would help assess practicality.

Questions for authors
  1. Can the authors compare GNI directly against deterministic clean-score ranking, keeping pseudo-group selection and the downstream training procedure identical? Also, if possible compare method against at least two recent group-free baselines (e.g., LFR, EVaLS, SAGE, Evidential Alignment) on atleast one bechmark.

  2. How sensitive are the results to the perturbation magnitude, number of views, pseudo-group selection thresholds, and reliability exponents?

  3. Can the authors provide component ablations on CelebA and Waterbirds to establish whether the individual contributions generalize beyond MIMIC-CXR+NIH?

  4. How does performance change if D-CORE is disabled at test time (with the same trained model)? Does D-CORE ever hurt mean/WGA materially on some samples?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 1 on Pseudo-Group Discovery and Ablation Studies

Major1&Minor1&Q1: We thank the reviewer for the questions. NSF’s sign-based grouping can fail under weak bias. To address this limitation, GNI uses intra-class ranking, which can correctly separate two groups under weaker bias conditions in our theoretical model. Using multiple perturbed views can further improve pseudo-group quality. Detailed proofs are provided in Appendix A. We added clean-score baseline that applies the same intra-class ranking to noise-free scores with all other settings held fixed (Appendix D, Table 4). We also varied the number of perturbed views to assess effects on pseudo-group quality and downstream performance (Appendix D, Table 5). Under our assumptions, the upper bound on the probability of reversing score order decreases exponentially with the number of independent views. Greater ranking stability can in turn improve pseudo-group assignment accuracy. We also added comparisons with EVaLS, Evidential Alignment. We omitted LFR because we found no public implementation. As reported in Appendix F, our method outperforms all two baselines in the evaluated settings. Major2&Q2&Q3&Q4: We extended the ablations to all seven settings (Section 4.3, Table 2). GNI with reliability weighting already performs well under weaker spurious correlations. Under stronger correlations, D-CORE complements training-stage debiasing by aggregating multi-layer bias-expert margin differences between factual features and zero-feature references into a sample-specific correction of the final decision margin. This correction incurs no loss in mean accuracy. We also added hyperparameter sensitivity analysis (Appendix D1&D8, Table5&Table12 ). Performance is sensitive to the voting threshold, which determines which samples are assigned to pseudo-groups and thus affects their quality. We acknowledge that original reliability ablation conflated component removal with changes in exponentiation. The revised consistency-only and rank-only configurations retain their respective exponents to assess complementary cross-view agreement and intra-class position. These results and the clean-score ablation are reported in Appendix D, Table 4. With their respective exponents retained, both the consistency-only and rank-only configurations outperform uniform weighting, where all reliability weights are set to one. We also conducted paired Wilcoxon signed-rank tests (Appendix F). Our method achieves statistically significant improvements in test-set worst-group accuracy (WGA) over each evaluated baseline. Major3: We thank the reviewer for the question about D-CORE. In Appendix B.1, we added analysis of residual correction under controlled interventions on spurious features. The results show that zero-feature responses provide an effective reference for isolating residual shortcut signals and that aggregating these signals to correct predictions improves prediction consistency under the interventions. These findings suggest that D-CORE can reduce the influence of spurious features. Minor3: We thank the reviewer for raising the concern about lower mean accuracy on some benchmarks, particularly Waterbirds. This behavior may be due to reduced reliance on spurious features. For example, on CelebA, ERM achieves the highest mean accuracy but the lowest WGA, whereas several debiasing methods achieve higher WGA with lower mean. One explanation is that reducing reliance on spurious features can improve predictions for the worst-performing group while leading to errors on some samples that were previously classified correctly using spurious features. These additional errors can in turn lead to a decrease in mean. Thus we interpret the lower mean in some settings as a possible consequence of our method reducing model’s reliance on spurious features. We thank the reviewer for indicating certain decreases in our mean. We will address this issue in future work. Minor4: We thank the reviewer for requesting clarification. For each method and dataset, we used the best configuration reported by the original authors and followed their threshold selection procedures. We selected checkpoints by validation WGA for most methods and retained GERNE’s official validation worst-class accuracy (WCA) criterion. Our method uses WCA, which requires only target labels, to accommodate scarce bias annotations in medical imaging. We also evaluated WGA-based checkpoint selection to assess group-informed validation. Appendix D.7, Table 11 shows that validation group annotations improve model selection and performance. Minor5: We thank the reviewer for asking about computational cost. In Appendix I, we compare total runtime and peak GPU memory usage with those of baselines. On MIMIC-CXR+NIH at bias ratio 0.90, our method takes longer to run but has the lowest peak GPU memory usage among the competing methods. It also achieves the best WGA among all competing methods. We will focus on reducing training time in future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Additional Analyses

Main Text: Secs 1-2. Clarified NSF's limitations under weak spurious correlations, strengthened the motivation for selective pseudo-group discovery, and expanded comparisons with recent debiasing and counterfactual methods to highlight our methodological distinctions. Sec 3. Refined mathematical formulations, standardized notation, and introduced NSF preliminaries. Clarified GNI-based ranking, selective assignment, reliability-weighted optimization, invariant prototypes, and D-CORE correction. Sec 4. Extended component ablations to all seven settings (Table 2), demonstrating the effectiveness of GNI-based training and complementary improvements from prototype regularization and D-CORE. Supplementary Material: Sec A.1. Established an idealized core-bias representation model with fixed covariance-aware geometry to analyze pseudo-group separation under varying bias strengths. Sec A.2. Derived the population score gap Δ=8χB², proving that intra-class ranking separates groups for χB²>0, relaxing NSF's requirement χB²>1. Sec A.3. Proved that Gaussian interventions preserve expected scores, reduce intervention-induced variance at O(1/V), and yield an exponentially decreasing ranking-reversal probability bound, theoretically explaining multi-view stability. Sec B. Formulated an additive residual-bias model and derived the variance-minimizing D-CORE coefficient, establishing sufficient conditions for reducing expected invariant-margin variance. Sec B.1. Conducted controlled Waterbirds background interventions. D-CORE reduced background-induced margin shifts and improved prediction consistency, supporting residual-shortcut mitigation. Sec B.2. Extended D-CORE to multiclass classification through centered-logit and pairwise-margin correction, theoretically establishing conditions for reducing multiclass spurious-feature dependence. Sec C.2. Clarified implementation details, baseline configurations, checkpoint selection, and validation protocols, emphasizing WCA-based model selection without requiring spurious-group annotations. Sec C.3. Specified correction-strength search and validation-based selection with a WCA constraint, ensuring consistent correction across test samples without test-label information. Sec D (Table 4). Disentangled GNI scoring, selective assignment, and reliability components. Joint weighting achieved 66.24% WGA versus 55.80% with uniform weighting, validating their complementary contributions. Sec D.1 (Table 5). Increasing intervention views from 2 to 32 improved WGA (58.80% to 67.80%) and conflict precision (87.79% to 99.24%). Eight views provide a practical accuracy-efficiency balance. Sec D.2 (Table 6). Assessed pseudo-group quality using held-out group annotations. Compared with NSF, conflict precision increased from 42.42% to 97.69%, confirming substantially improved assignment reliability. Sec D.3 (Table 7). Compared fixed and validation-selected thresholds. Validation selection improved WGA by 1.06-4.96 points. Sec D.4 (Table 8). Adapted CF-VQA using its official implementation under matched training settings. D-CORE achieved higher WGA and Mean across all seven settings, validating multi-level residual correction. Sec D.5 (Table 9). Ablated invariant-prototype regularization while fixing other components. Consistent WGA improvements across seven settings confirm the benefit of group-balanced prototype alignment. Sec D.6 (Table 10). Compared Euclidean and regularized Mahalanobis metrics. The latter improved WGA and Mean across seven settings, including Waterbirds WGA from 73.77% to 88.19%. Sec D.7 (Table 11). WGA-based checkpoint selection improved test WGA by 0.69-4.91 points over WCA-based selection, quantifying the benefit of group-informed validation. Sec D.8 (Table 12). Analyzed perturbation magnitude, voting threshold, and reliability exponents. Voting threshold showed the greatest sensitivity, highlighting the coverage-reliability trade-off. Sec E (Table 13). Added comparisons with EVaLS and Evidential Alignment. Our method achieved the highest WGA across all seven settings and highest Mean in six, demonstrating superior robustness. Sec F (Table 14). Paired Wilcoxon tests over five seeds confirmed statistically significant WGA improvements over six baselines across seven settings (p<0.05). Sec G (Table 15). Evaluated simultaneous age and gender shortcuts on HAM10000. WGA improved from 61.28% to 64.52%, and Mean from 77.05% to 81.38%, demonstrating practical multi-shortcut robustness. Sec H (Table 16). Evaluated unbiased Waterbirds with balanced spurious groups. WGA improved from 87.95% to 92.31% and Mean from 92.03% to 93.68%, demonstrating robustness without dominant shortcuts. Sec I (Table 17). Compared runtime and peak GPU memory with six baselines. Despite longer runtime (8420.69 s), our method achieved the lowest peak memory (2036 MB), highlighting its memory efficiency.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.