acceptodds
Under review as a conference paper at ICLR 2027

IsoAct: Geometry-Aware Post-hoc Debiasing via Isometric Actions

Abstract

Learning representations that are robust to nuisance factors is essential for reliable deployment, but remains challenging when the representations lie on non-Euclidean manifolds. Existing debiasing methods typically assume Euclidean representations or apply linear interventions, which can distort the geometry of structured latent spaces. We propose IsoAct, a post-hoc framework that accounts for geometry in both nuisance identification and representation editing. It identifies nuisance-predictive directions using Fisher discriminant analysis (FDA) in tangent coordinates and applies corrections through manifold-native actions that preserve manifold membership. Controlled and real-world experiments on hyperspherical, hyperbolic, and representations evaluate nuisance reduction, utility preservation, and manifold validity jointly. IsoAct reduces nuisance recoverability while retaining competitive utility. Its clearest joint result is on KLIFS, combining reduced kinase-group recoverability, strong structural prediction, and valid rigid poses.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Fairness and Debiasing, Nuisance-Robust Representations, Post-hoc Concept Erasure, Manifold-aware Representation Learning

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_5gmBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes IsoAct, a post-hoc debiasing framework for representations on known manifolds, including the hypersphere, Lorentz hyperboloid, and SE(3). It identifies nuisance-predictive directions using Fisher Discriminant Analysis (FDA) in tangent coordinates, constructs a neutral reference from low-confidence training examples, and applies manifold-native updates to reduce nuisance predictability while preserving manifold membership. Experiments on synthetic benchmarks, KLIFS ligand poses, and CLIP/MERU embeddings evaluate nuisance reduction, downstream utility, and geometric validity.

Critical strengths
  1. Debiasing manifold-valued representations is a meaningful and relatively underexplored problem. Jointly evaluating nuisance reduction, downstream utility, and manifold validity provides a useful perspective.

  2. The method offers a coherent post-hoc pipeline across three distinct geometries, with reasonably detailed mathematical formulations and implementation specifications.

  3. The empirical results are encouraging in several settings, particularly the consistent debiasing improvements on MERU and the high retrieval retention on CLIP.

Critical weaknesses
  1. IsoAct provides no guarantee of linear nuisance erasure. Matching projections along FDA directions does not necessarily equalize class means, so the nuisance can remain recoverable by linear probes, unlike LEACE’s guaranteed linear erasure in Euclidean space.

  2. The real-world experiments do not isolate the benefits of manifold-native editing: KLIFS selects translation-only updates, all MERU baselines are already manifold-valid, and CLIP lacks comparisons with renormalized baselines. Natural competitors such as Euclidean editing with projection, log–LEACE–exp, and WRING are also missing.

Questions for authors
  1. How does IsoAct compare with Euclidean editing followed by manifold projection, or with a log–LEACE–exp baseline?

  2. Why does the selection rule pick translation-only editing on KLIFS, and how much kinase-group information remains in the unedited rotations (e.g., from a probe on rotations alone)?

4 comments

Thank you for recognizing the importance of manifold-valued debiasing, the coherent framework across three geometries, and the encouraging MERU and CLIP results. Your suggestions on projection baselines and component-wise probes guided the additional KLIFS experiments below.

1. Euclidean editing followed by projection (Q1/W2). By evaluating the performance of 'projected' variants of baselines in KLIFS setting, we found that projection restores manifold constraints, but its effects on debiasing and utility are not coherent.

The details of the projection are as follow: we project each selected output rotation onto SO(3), with determinant correction. Raw/projected pairs share the debiasing configuration. Entries of the table are full-test estimates ± SD from 1,000 paired test bootstrap resamples with fitted models fixed. C1/C2 are mean orthogonality/determinant errors.

Also, we have added LEACE as a new baseline to compare their empirical debiasing–utility trade-offs. While your suggestion of 'log–LEACE–exp' method is also valid, we chose to apply the same post-projection repair protocol to all baselines for consistency.

N.Acc and N.F1 denote kinase-group prediction accuracy and macro-F1. aC and Pocket are the two downstream tasks; their F1 and Rec scores concern the minority class, and AUC denotes AUROC. Null is the shuffled-label nuisance reference: proximity to Null indicates stronger debiasing, while higher utility scores are better. C1/C2 measure rotation orthogonality/determinant errors, for which lower is better.

MethodN.AccN.F1aC F1aC RecaC AUCPocket F1Pocket RecPocket AUCC1C2
Original0.1398±0.00670.1409±0.00670.3836±0.02460.3362±0.02520.7163±0.01510.7929±0.01750.7357±0.02300.9708±0.00206.5065e-082.2077e-08
Null0.0694±0.00220.0603±0.0021————————
CMF0.1988±0.00800.1858±0.00740.3464±0.02380.3046±0.02410.7087±0.01560.8273±0.01190.8173±0.01260.9732±0.00203.6900e-011.1368e-01
CMF+Proj0.1645±0.00730.1645±0.00720.3744±0.02410.3190±0.02420.7149±0.01580.7961±0.01780.7535±0.02260.9720±0.00211.1711e-155.8252e-16
SFID0.1398±0.00670.1409±0.00670.3717±0.02500.3247±0.02540.7176±0.01490.7867±0.01740.7388±0.02260.9712±0.00216.5065e-082.2077e-08
SFID+Proj0.1398±0.00670.1409±0.00670.3606±0.02510.3103±0.02500.7183±0.01490.7863±0.01740.7384±0.02260.9712±0.00211.1095e-155.6016e-16
SPD0.1425±0.00680.1415±0.00670.2862±0.02390.2414±0.02240.7061±0.01440.8233±0.01430.7797±0.01890.9729±0.00199.4948e-013.7450e-01
SPD+Proj0.1699±0.00750.1601±0.00720.2527±0.02410.2040±0.02160.6855±0.01470.8454±0.01140.8182±0.01370.9730±0.00211.1599e-155.9305e-16
LEACE0.1306±0.00630.0647±0.00360.3746±0.02530.3046±0.02440.6842±0.01630.7962±0.01300.7438±0.01630.9545±0.00321.0036e+006.5431e-01
LEACE+Proj0.1202±0.00650.0925±0.00480.3122±0.02510.2615±0.02390.7200±0.01360.7274±0.01980.6519±0.02370.9357±0.00441.1151e-155.6640e-16
IsoAct0.1117±0.00620.1022±0.00600.4277±0.02360.3822±0.02480.7333±0.01590.8084±0.01460.8174±0.01430.9664±0.00246.9510e-163.5723e-16

The results show why geometric validity alone is insufficient. Projection has mixed effects on debiasing and utility. For SPD, projection repair worsens debiasing while showing mixed impact on utility preservation. For LEACE, projection brings nuisance Acc closer to Null, yet decreases five of six utility metrics. IsoAct combines reduced nuisance predictability with the highest aC F1, Recall, and AUROC in this comparison and near-zero constraint errors. Projection repair alone does not reproduce this trade-off.

2. Kinase-group information in rotation and translation component (Q2/W2). To examine the bias information in each component, we used the same split and train-fitted linear probe:

R (9D), t (3D), and R+t (12D) denote the flattened rotation, translation, and their concatenation, respectively. The same nuisance metrics and bootstrap SD protocol apply.

FeaturesN.Acc±SDN.F1±SD
Null0.0694±0.00220.0603±0.0021
Original R (9D)0.1036±0.00610.0879±0.0054
Original t (3D)0.1560±0.00720.1486±0.0067
Original R+t (12D)0.1398±0.00670.1409±0.0067

Although the rotation component also contains probe-readable kinase-group information, translation supports stronger prediction. This provides an empirical explanation for why translation-only editing was selected in this setting.

What's more, we would like to clarify that the selection rule did not specify the type of IsoAct configurations. The selection rule balances nuisance reduction and utility across available actions, and the translation-only editing was chosen among candidate pool.

3. Linear nuisance erasure (W1). We agree that IsoAct does not provide LEACE's guarantee of linear nuisance erasure. Matching projections along FDA directions does not by itself guarantee equal nuisance-class means, so linear predictability may remain. IsoAct aims to reduce nuisance predictability while retaining utility and manifold membership. We therefore added LEACE and its projected variant to compare these trade-offs empirically, as shown above.

We will add these comparisons and component-wise probes to the revision to clarify both the effect of projection and the selection of translation-only editing.

I appreciate the authors’ additional experiments and clarification. The component-wise probes explain the translation-only choice on KLIFS, and the projection-repair comparisons strengthen the evaluation. However, several key concerns remain: The CLIP experiments still lack SPD with L2 normalization and WRING, which is directly applicable to this setting. For KLIFS, the main result still relies on translation-only editing. The joint-editing experiment demonstrates feasibility but does not isolate the contribution of SO(3) editing. So the experiments do not establish a clear advantage over the simplest or closely related alternatives. Therefore, I will maintain my score.

Thank you for recognizing that the component-wise probes explain the translation-only choice and that the projection-repair comparisons strengthen the evaluation. We present further CLIP comparisons and a controlled KLIFS ablation to address your remaining concerns.

1. SPD with L2 normalization and WRING on CLIP.

IsoAct achieves higher mean retrieval retention than SPD+Norm and WRING across all three attributes, with the highest DD for age.

Under the same setting as our reported CLIP results, we added SPD followed by L2 normalization (SPD+Norm) and WRING. Following Table 3 of the revised manuscript, DD averages

1−∣mmethod−mNull∣∣mOriginal−mNull∣1-\frac{|m_{\mathrm{method}}-m_{\mathrm{Null}}|}{|m_{\mathrm{Original}}-m_{\mathrm{Null}}|}

over protected-attribute Accuracy and macro-F1. Higher DD indicates greater proximity to Null, the shuffled-label reference. Retrieval retention averages the method/Original ratios for R@1, R@5, and R@10. The constraint column reports mean unit-norm error. Bold marks the highest DD and retention within each attribute among the reported methods.

AttributeMethodMean DD ↑Mean retrieval retention ↑Mean unit-norm error ↓
AgeSPD0.75360.98671.8327e-02
AgeSPD+Norm0.81110.98683.3660e-09
AgeWRING0.79840.98812.2936e-08
AgeIsoAct0.92340.98906.3230e-08
GenderSPD0.82430.99381.0768e-02
GenderSPD+Norm0.42890.99393.1858e-09
GenderWRING0.37040.99902.3162e-08
GenderIsoAct0.42541.00033.5303e-08
RaceSPD0.96560.98915.3183e-02
RaceSPD+Norm0.93240.99073.5094e-09
RaceWRING0.88100.99262.2442e-08
RaceIsoAct0.80220.99326.1190e-08

Age. Normalization improves SPD’s debiasing, while IsoAct achieves the highest DD and retrieval retention among the compared methods.

Gender. IsoAct exceeds WRING on both metrics and achieves higher retrieval retention than SPD+Norm at a similar DD. Normalization reduces SPD’s DD from 0.8243 to 0.4289 despite restoring near-zero unit-norm error.

Race. SPD+Norm and WRING achieve higher DD, while IsoAct achieves higher mean retrieval retention.

These comparisons address the simpler normalization baseline and directly applicable WRING method you identified as missing. Since SPD+Norm, WRING, and IsoAct all achieve near-zero unit-norm error, these results strengthen the empirical case for IsoAct beyond satisfying the output constraint.

2. Isolating the contribution of rotational editing on KLIFS.

To address your concern that the joint-editing experiment "does not isolate the contribution of SO(3) editing," we conducted a controlled ablation.

Starting from the joint configuration in our previous response, we restored the original rotation of each edited pose (RJ,tJ)(R_J,t_J) to obtain (R,tJ)(R,t_J). The two conditions differ only in rotation; translation outputs are exactly identical, and all other experimental settings are unchanged.

N.Acc and N.macro-F1 measure kinase-group prediction. Both conditions remain above Null, so lower nuisance scores are better here; higher task scores are better. Entries are full-test estimates ± SD from 1,000 paired test bootstrap resamples. Bold marks the better point estimate.

ConfigurationN.AccN.macro-F1aC F1aC RecallaC AUROCPocket F1Pocket RecallPocket AUROC
Original rotations restored0.1314 ± 0.00660.1302 ± 0.00660.3907 ± 0.02510.3391 ± 0.02550.7262 ± 0.01480.8355 ± 0.01150.8203 ± 0.01140.9717 ± 0.0021
Joint editing0.1283 ± 0.00670.1167 ± 0.00630.3922 ± 0.02480.3448 ± 0.02530.7382 ± 0.01410.8288 ± 0.01290.8241 ± 0.01220.9722 ± 0.0022

Keeping the edited rotations brings both nuisance scores closer to Null, with macro-F1 decreasing from 0.1302 to 0.1167. It also yields higher point estimates for all three aC metrics and Pocket Recall/AUROC; Pocket F1 is lower (0.8288 versus 0.8355).

Thus, this controlled comparison isolates a contribution of SO(3) editing to nuisance reduction and downstream performance with translation held fixed.

We sincerely appreciate the time and attention you have devoted to our paper throughout the discussion. Your feedback has helped us strengthen the empirical evaluation and clarify the scope and contributions of our work.

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

The paper proposes IsoAct, a geometry-aware post-hoc debiasing framework for representations on non-Euclidean manifolds. It identifies nuisance-predictive directions using Fisher Discriminant Analysis (FDA) in tangent coordinates, constructs a neutral target from low-confidence training examples, and applies manifold-native transformations that preserve manifold membership. The framework supports hyperspherical, Lorentz-hyperbolic, and SE(3) representations. Experiments on synthetic datasets, KLIFS, CLIP, and MERU demonstrate promising nuisance reduction while maintaining downstream utility and geometric validity.

Critical strengths
  1. Well-motivated and relevant problem. The paper addresses an important limitation of conventional post-hoc debiasing methods: Euclidean feature modifications may violate the geometric constraints of non-Euclidean representations.
  2. Unified framework across different manifolds.hyperspheres, Lorentz hyperboloids, and SE(3).
Critical weaknesses

Weaknesses

  1. Missing ablation studies. The paper does not isolate the contributions of tangent-space FDA, the low-confidence neutral target, and manifold-native actions.

  2. Limited geometric distortion analysis. Theorem 4.1 guarantees manifold membership but not pairwise-distance preservation. The paper does not analyze how such distortions relate to downstream-task performance.

  3. Lack of multi-seed evaluation. Results are reported without standard deviations or confidence intervals across independent runs, limiting assessment of statistical reliability.

  4. Limited comparison with related methods. Despite discussing WRING (Gerych et al., 2026), the paper lacks a direct comparison on shared benchmarks, making it difficult to assess the advantages of IsoAct over existing rotation-based debiasing methods.

Questions for authors
  1. Tangent-space FDA: How does it compare with ambient-space FDA using the same correction actions?

  2. Neutral target: How does the low-confidence target compare with the global mean, and how sensitive is it to the confidence threshold?

  3. Manifold-native actions: How does IsoAct compare with Euclidean debiasing followed by manifold projection?

  4. Geometric distortion: How do changes in pairwise distances affect downstream performance?

  5. Statistical robustness: Can the authors report mean and standard deviation across multiple random seeds?

4 comments

Thank you for recognizing the motivation for geometry-aware debiasing and the unified treatment of three manifolds. Your suggestions guided additional KLIFS experiments to isolate the method's components and examine geometric distortion and statistical reliability.

1. Tangent-space FDA (Q1/W1). We compared tangent and ambient FDA while keeping other settings identical and changing only the FDA learning space.

N.Acc/F1 denote kinase-group accuracy/macro-F1; aC and Pocket denote the downstream tasks. Task F1/Recall concern the minority class. Entries are full-test estimates ± SD from 1,000 paired bootstrap resamples with fitted models fixed.

FDAN.AccN.F1aC F1aC RecallaC AUROCPocket F1Pocket RecallPocket AUROC
Tangent FDA0.1117 ± 0.00620.1022 ± 0.00600.4277 ± 0.02360.3822 ± 0.02480.7333 ± 0.01590.8084 ± 0.01460.8174 ± 0.01430.9664 ± 0.0024
Ambient FDA0.1337 ± 0.00660.1298 ± 0.00640.4048 ± 0.02490.3420 ± 0.02520.7477 ± 0.01430.8059 ± 0.01590.7745 ± 0.02000.9707 ± 0.0021

While ambient FDA shows better performance than tangent FDA in some utility metrics, it shows much weaker debiasing ability. This supports using tangent FDA for better nuisance reduction with utility preservation.

2. Neutral target (Q2/W1). We tested two aspects of the reported configuration:

  • Low-confidence versus global mean: replacing the target at τ=.7\tau=.7 with the global training mean
  • Threshold sensitivity: for τ∈{.5,.6,.7,.8,.9}\tau\in\{.5,.6,.7,.8,.9\}

The ablation experiments show that the performance is stable across the tested targets and thresholds. Also, we would like to clarify that the default setting of τ=.7\tau=.7 follows the prior work SFID.

TargetN.AccN.F1
Global mean0.1113±0.00610.1019±0.0059
Low confidence, τ=0.5\tau=0.50.1117±0.00620.1022±0.0060
Low confidence, τ=0.6\tau=0.60.1113±0.00610.1019±0.0059
Low confidence, τ=0.7\tau=0.7 (reported)0.1117±0.00620.1022±0.0060
Low confidence, τ=0.8\tau=0.80.1113±0.00610.1019±0.0059
Low confidence, τ=0.9\tau=0.90.1117±0.00620.1022±0.0060

All six utility metrics were unchanged across these target choices, and both mean constraint residuals were identical: C1=6.9510×10−166.9510\times10^{-16} and C2=3.5723×10−163.5723\times10^{-16}.

3. Editing plus projection (Q3; W1/W4). Projection restores manifold constraints, but its effects on debiasing and utility vary. We project each selected rotation onto SO(3), and raw/projected pairs share configurations. Entries of the table are test estimates ± bootstrap SD (Q5).

Furthermore, regarding the concern of limited comparison with related methods, LEACE is added as a new baseline. While WRING is relevant to IsoAct, its design is confined to using Euclidean norms and extending it to preserve hyperbolic or SE(3) requires additional geometric design. Thus, only LEACE was added as a new baseline.

The results show that projection repair alone does not reproduce IsoAct's debiasing–utility trade-off. While all mean orthogonality/determinant errors of all projected variants are below 1.2×10−151.2\times10^{-15}, projection shows inconsistent impact on performances. For SPD, projection repair worsens debiasing while showing mixed impact on utility preservation. For LEACE, projection improves debiasing ability, yet decreases five of six utility metrics. IsoAct achieves the highest aC F1, Recall, and AUROC in this comparison while reducing nuisance predictability and preserving manifold validity, supporting the value of manifold-native editing beyond projection repair.

C1/C2 are mean rotation orthogonality/determinant errors; lower is better.

MethodN.AccN.F1aC F1aC RecaC AUCPocket F1Pocket RecPocket AUCC1C2
Original.1398±.0067.1409±.0067.3836±.0246.3362±.0252.7163±.0151.7929±.0175.7357±.0230.9708±.00206.5065e-082.2077e-08
Null.0694±.0022.0603±.0021————————
CMF.1988±.0080.1858±.0074.3464±.0238.3046±.0241.7087±.0156.8273±.0119.8173±.0126.9732±.00203.6900e-011.1368e-01
CMF+Proj.1645±.0073.1645±.0072.3744±.0241.3190±.0242.7149±.0158.7961±.0178.7535±.0226.9720±.00211.1711e-155.8252e-16
SFID.1398±.0067.1409±.0067.3717±.0250.3247±.0254.7176±.0149.7867±.0174.7388±.0226.9712±.00216.5065e-082.2077e-08
SFID+Proj.1398±.0067.1409±.0067.3606±.0251.3103±.0250.7183±.0149.7863±.0174.7384±.0226.9712±.00211.1095e-155.6016e-16
SPD.1425±.0068.1415±.0067.2862±.0239.2414±.0224.7061±.0144.8233±.0143.7797±.0189.9729±.00199.4948e-013.7450e-01
SPD+Proj.1699±.0075.1601±.0072.2527±.0241.2040±.0216.6855±.0147.8454±.0114.8182±.0137.9730±.00211.1599e-155.9305e-16
LEACE.1306±.0063.0647±.0036.3746±.0253.3046±.0244.6842±.0163.7962±.0130.7438±.0163.9545±.00321.0036e+006.5431e-01
LEACE+Proj.1202±.0065.0925±.0048.3122±.0251.2615±.0239.7200±.0136.7274±.0198.6519±.0237.9357±.00441.1151e-155.6640e-16
IsoAct.1117±.0062.1022±.0060.4277±.0236.3822±.0248.7333±.0159.8084±.0146.8174±.0143.9664±.00246.9510e-163.5723e-16

4. Distance changes and utility (Q4/W2). IsoAct's sample-dependent actions preserve membership without requiring pairwise-distance preservation. We measured all 3,368,310 test pairs using Euclidean distances on native-unit [vec(R),t][\mathrm{vec}(R),t]. Normalized distortion is

D=∑i<j∣dij′−dij∣∑i<jdijD = \frac{\sum_{i < j} \lvert d_{ij}^{\prime} - d_{ij} \rvert}{\sum_{i < j} d_{ij}}

Here dd/d′d' denote Original/edited distances. Methods share the denominator; F1 changes are relative to Original. Mean pairwise distance is the average distance between each method's test representations over the same pair set.

MethodMean pairwise distanceDistortionΔaC F1ΔPocket F1
Original4.66030+.0000+.0000
CMF4.62878.4076e-03−.0372+.0345
SFID4.66031.6170e-07−.0119−.0062
SPD3.72032.0169e-01−.0974+.0304
LEACE3.33123.3729e-01−.0090+.0033
IsoAct2.48184.7846e-01+.0440+.0155

IsoAct reduces mean pairwise distance from 4.6603 to 2.4818 and shows the largest distance distortion in this comparison, yet improves both task F1 scores. SFID changes pairwise distances very little but slightly lowers both F1 scores. Thus, greater distance distortion does not necessarily imply worse downstream performance.

5. Statistical robustness (Q5/W3). Following the classification evaluation protocol of the prior work SPD, we report full-test estimates ± SD from 1,000 bootstrap resamples. These quantify test-sample uncertainty.

We will include these ablations and comparisons in the revision.

Reviewer_EFCCOpenReview ↗

Concerns Addressed

Many of my concerns are addressed; thus, I am inclined to raise my score. However, due to my low familiarity with this field, I request the AC to provide less weightage for my score.

Thank you for your thoughtful feedback and for taking the time to review our responses. We are glad that the additional experiments and clarifications have addressed many of your concerns. Your comments have helped us strengthen the evaluation and improve the paper.

Reviewer_YkX1OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes IsoAct, a post-hoc debiasing framework for manifold-valued representations. It identifies nuisance-predictive directions using Fisher discriminant analysis (FDA) in tangent coordinates, constructs a neutral reference from low-confidence samples, and applies geometry-specific editing operations on hyperspherical, hyperbolic, and SE(3) representations. The paper establishes a manifold-membership guarantee and evaluates nuisance reduction, downstream utility, and geometric validity on synthetic benchmarks, KLIFS ligand poses, and CLIP/MERU embeddings.

Critical strengths
  1. The paper addresses an important limitation of conventional post-hoc debiasing, where unconstrained feature editing can violate the geometry of structured representations. The proposed operations preserve manifold membership by construction, and the theoretical analysis adequately supports this guarantee.

  2. IsoAct demonstrates reduced nuisance predictability while retaining competitive downstream utility and valid representations. The KLIFS results show improvements in several structural prediction metrics alongside reduced nuisance recoverability, while the CLIP and MERU experiments further demonstrate the applicability of the framework.

Critical weaknesses
  1. The core approach combines existing ingredients, including tangent-space discriminant analysis, logarithmic maps, parallel transport, and geometry-preserving transformations. Prior work on Riemannian classification (such as Barachant et al. 2012), has explored tangent-space discriminant learning, while related debiasing methods have investigated geometry-aware interventions. The three manifold instances rely on separately designed operations rather than a substantially new general construction, and the manifold-membership guarantee largely follows from standard geometric properties.

  2. The experiments do not adequately isolate the benefits of manifold-native editing from conventional debiasing followed by geometric correction. In MERU, all methods satisfy the Lorentz constraint through coordinate reconstruction, whereas the Euclidean baselines on CLIP and SE(3) are not evaluated with analogous normalization or rotation projection. In particular, SPD achieves stronger CLIP debiasing on some attributes with comparable retrieval utility, but its performance after unit-norm normalization remains unknown.

  3. The selected KLIFS configuration uses translation-only actions, which preserve SE(3) membership without requiring rotational editing. Therefore, this result does not directly substantiate the effectiveness of the proposed SO(3) action. Moreover, the reported rotation-constraint residuals improve from approximately 10−810^{-8} to 10−1610^{-16} despite the translation-only configuration, suggesting an additional stabilization or projection step. The absence of a comparably constrained translation-only baseline further limits the evidence that the observed improvements arise from IsoAct's specific geometric construction.

  4. Although the paper discusses WRING and LEACE, neither is included in the empirical comparisons. WRING is particularly relevant because it also replaces projection-based debiasing with rotational interventions, while LEACE provides an established linear concept-erasure baseline. Without comparisons to these closely related approaches, it remains difficult to determine whether IsoAct offers a meaningful advantage over existing methods beyond preserving manifold validity.

Barachant, A., Bonnet, S., Congedo, M., and Jutten, C. (2012). Multiclass brain-computer interface classification by Riemannian geometry. IEEE Transactions on Biomedical Engineering, 59(4), 920–928. https://doi.org/10.1109/TBME.2011.2172210

Questions for authors
  1. What is the key technical contribution beyond combining tangent-space FDA with existing geometry-preserving transformations? What new principle, capability, or property distinguishes IsoAct from established geometric representation-editing methods?

  2. How does IsoAct compare with conventional debiasing followed by appropriate manifold projection, particularly SPD with unit-norm normalization on CLIP and SO(3) projection for SE(3) representations? Could such baselines achieve comparable nuisance reduction and downstream utility while satisfying the same geometric constraints?

  3. Since the selected KLIFS configuration uses translation-only actions, how does IsoAct compare with a simpler translation-only debiasing baseline under the same constraints? What evidence establishes the benefit of the proposed geometry-guided correction rather than merely restricting edits to translation?

  4. Does the translation-only KLIFS configuration involve an additional SO(3) projection or stabilization step, and does this explain the improvement in rotation-constraint residuals? Beyond the synthetic experiments, what evidence supports the practical contribution of the proposed rotational editing mechanism?

  5. How does IsoAct compare empirically with closely related approaches such as WRING and LEACE under comparable settings? If direct comparisons are not feasible, what methodological differences make these baselines unsuitable for evaluating the claimed advantages of IsoAct?

4 comments

Thank you for recognizing the manifold-membership guarantee and the empirical results on KLIFS, CLIP, and MERU. Your comments on prior work, projection baselines, and rotational editing helped us clarify the contribution and guided the additional comparisons below.

1. Technical contribution (Q1/W1). IsoAct provides a framework that connects manifold-aware nuisance directions to manifold native actions for debiasing. Adjusting to each manifolds is needed as isometric actions are defined differently, yet IsoAct still implements the same principle: identify nuisance directions, compute corrections, and apply native actions. Tangent FDA identifies directions; IsoAct specifies how to use them and how to edit the original representation on its manifold.

The distinction from Barachant et al. (2012) is the role of the learned directions. Their TSLDA method maps covariance matrices to a Riemannian tangent space, selects discriminative variables, and trains an LDA classifier. IsoAct uses nuisance-predictive directions to modify an existing representation toward a neutral target through sample-specific native actions. Learning a discriminant direction alone does not specify this correction or how to realize it on the original manifold. The manifold-specific formulas implement this shared editing procedure using the actions available in each geometry.

The contribution is this direction-to-action editing procedure across geometries; the comparisons below test its benefits beyond repairing Euclidean edits.

2. Projection and related methods (Q2/Q5; W2/W4). We compared each baseline with the same projection repair: project edited rotations onto SO(3) with determinant correction. Raw/projected pairs share configurations.

We also added LEACE as a new baseline. While WRING is relevant, it uses Euclidean norms and cosine similarity. Extending it to preserve hyperbolic or SE(3) constraints requires additional geometric design, so only LEACE was added as a new baseline.

N.Acc/F1 measure nuisance prediction; Null is a shuffled-label reference. Rec/AUC mean Recall/AUROC. Entries are test estimates ± SD from 1,000 paired bootstrap resamples with models fixed.

MethodN.AccN.F1aC F1aC RecaC AUCPocket F1Pocket RecPocket AUC
Original.1398±.0067.1409±.0067.3836±.0246.3362±.0252.7163±.0151.7929±.0175.7357±.0230.9708±.0020
Null.0694±.0022.0603±.0021——————
CMF.1988±.0080.1858±.0074.3464±.0238.3046±.0241.7087±.0156.8273±.0119.8173±.0126.9732±.0020
CMF+Proj.1645±.0073.1645±.0072.3744±.0241.3190±.0242.7149±.0158.7961±.0178.7535±.0226.9720±.0021
SFID.1398±.0067.1409±.0067.3717±.0250.3247±.0254.7176±.0149.7867±.0174.7388±.0226.9712±.0021
SFID+Proj.1398±.0067.1409±.0067.3606±.0251.3103±.0250.7183±.0149.7863±.0174.7384±.0226.9712±.0021
SPD.1425±.0068.1415±.0067.2862±.0239.2414±.0224.7061±.0144.8233±.0143.7797±.0189.9729±.0019
SPD+Proj.1699±.0075.1601±.0072.2527±.0241.2040±.0216.6855±.0147.8454±.0114.8182±.0137.9730±.0021
LEACE.1306±.0063.0647±.0036.3746±.0253.3046±.0244.6842±.0163.7962±.0130.7438±.0163.9545±.0032
LEACE+Proj.1202±.0065.0925±.0048.3122±.0251.2615±.0239.7200±.0136.7274±.0198.6519±.0237.9357±.0044
IsoAct T-only.1117±.0062.1022±.0060.4277±.0236.3822±.0248.7333±.0159.8084±.0146.8174±.0143.9664±.0024

The results show that restoring geometric validity does not by itself reproduce IsoAct's debiasing–utility trade-off. All projected variants achieve mean orthogonality/determinant errors below 1.2×10−151.2\times10^{-15}, yet SPD+Proj moves both nuisance scores farther from Null and lowers aC F1, while improving Pocket F1. LEACE+Proj brings nuisance Acc closer to Null but lowers five of six utility metrics. IsoAct achieves the highest aC F1, Recall, and AUROC in this comparison, improves both task F1 scores over Original, and maintains valid representations. These results support the value of directing the edits through native actions, beyond making an edited output geometrically valid.

3. Translation-only and joint editing (Q3/Q4; W3). Three observations address this:

First, the reported SFID configuration also changes only translation. SFID selects individual coordinates by their random-forest importance for nuisance prediction and replaces their values with low-confidence means. In our 12-dimensional KLIFS implementation, the number of edited coordinates is three. The training-based importance ranking selects all three translation coordinates; it is not constrained to select translation. While for both SFID and IsoAct changes are confined to translation component, IsoAct improves on its nuisance and task F1 scores, so restricting edits to translation alone does not explain the results.

Second, linear probes using the same split show stronger kinase-group prediction from translation:

FeaturesN.Acc±SDN.F1±SD
R (9D).1036±.0061.0879±.0054
t (3D).1560±.0072.1486±.0067
R+t (12D).1398±.0067.1409±.0067

This helps explain the choice. The selection rule balances nuisance reduction and utility across actions, without requiring translation-only editing.

Third, IsoAct configuration that edits both rotation and translation still achieves favorable results. Followings are performance of IsoAct when we constrain the candidate pool to exclude models that debiase only translation or rotation:

MethodN.AccN.F1aC F1aC RecaC AUCPocket F1Pocket RecPocket AUC
Original.1398.1409.3836.3362.7163.7929.7357.9708
Null.0694.0603——————
CMF.1988.1858.3464.3046.7087.8273‾\underline{.8273}.8173‾\underline{.8173}.9732
SFID.1398.1409.3717.3247‾\underline{.3247}.7176‾\underline{.7176}.7867.7388.9712
SPD.1425.1415.2862.2414.7061.8233.7797.9729‾\underline{.9729}
LEACE.1306‾\underline{.1306}.0647.3746‾\underline{.3746}.3046.6842.7962.7438.9545
IsoAct.1283.1167‾\underline{.1167}.3922.3448.7382.8288.8241.9722

Against raw CMF/SFID/SPD/LEACE, this configuration leads on all aC metrics, Pocket F1/Recall, and nuisance Acc proximity to Null. Thus, favorable performance also extends to joint editing.

4. Rotation residuals and the role of rotational editing (Q4/W3). The translation-only output has no additional SO(3) projection or stabilization: its rotation matrices are identical to the native input rotations. The reported residual difference comes from storing Original rotations in float32 and IsoAct outputs in float64; the residual calculations themselves use float64 in both cases. Using the same stored rotation values gives identical residuals, as shown below. Mean orthogonality/determinant errors:

Rotation valuesOrthog.Det.
Original, native float646.9510e-163.5723e-16
IsoAct, float646.9510e-163.5723e-16
Original, stored float326.5065e-082.2077e-08
IsoAct, rounded to float326.5065e-082.2077e-08

The joint KLIFS comparison above provides evidence beyond the synthetic experiments: a configuration that intervenes on rotation and translation retains a favorable trade-off against the raw baselines. Translation carries stronger probe-readable kinase-group information in KLIFS, which helps explain why translation-only editing was selected for the main table. Retaining rotational actions allows the framework to address nuisance information carried by rotation as well, rather than restricting it to cases where translation is the useful intervention.

Thanks for the detailed response and for some of the changes to the current manuscript.

The author’s explanations and additional experiments have, to a large extent, resolved my confusion regarding points W3, W4, Q3, Q4 and Q5 to the paper. The revised manuscript substantially strengthens the KLIFS evaluation through projection-repair comparisons and controlled tangent-versus-ambient FDA ablations.

However, the CLIP experiments still lack the requested unit-normalized SPD baseline and a direct comparison with the closely related rotation-based WRING method. Furthermore, the joint SE(3) results demonstrate feasibility rather than the independent benefit of rotational editing.

The remaining issues limit the strength of the claimed advantages. Consequently, I view this paper still falls short of the level of novelty and rigor required for the ICLR main track. I will adjust the overall rating to a score that corresponds to a "weak rejection."

Thank you for carefully reading our response and acknowledging the strengthened KLIFS evaluation. We present further CLIP comparisons and a controlled KLIFS ablation to address your remaining concerns.

1. Unit-normalized SPD and a direct comparison with WRING on CLIP.

IsoAct achieves higher mean retrieval retention than SPD+Norm and WRING across all three attributes, with the highest DD for age.

Under the same setting as our reported CLIP results, we added SPD followed by L2 normalization (SPD+Norm) and WRING. Following Table 3 of the revised manuscript, DD averages

1−∣mmethod−mNull∣∣mOriginal−mNull∣1-\frac{|m_{\mathrm{method}}-m_{\mathrm{Null}}|}{|m_{\mathrm{Original}}-m_{\mathrm{Null}}|}

over protected-attribute Accuracy and macro-F1. Higher DD indicates greater proximity to Null, the shuffled-label reference. Retrieval retention averages the method/Original ratios for R@1, R@5, and R@10. The constraint column reports mean unit-norm error. Bold marks the highest DD and retention within each attribute among the reported methods.

AttributeMethodMean DD ↑Mean retrieval retention ↑Mean unit-norm error ↓
AgeSPD0.75360.98671.8327e-02
AgeSPD+Norm0.81110.98683.3660e-09
AgeWRING0.79840.98812.2936e-08
AgeIsoAct0.92340.98906.3230e-08
GenderSPD0.82430.99381.0768e-02
GenderSPD+Norm0.42890.99393.1858e-09
GenderWRING0.37040.99902.3162e-08
GenderIsoAct0.42541.00033.5303e-08
RaceSPD0.96560.98915.3183e-02
RaceSPD+Norm0.93240.99073.5094e-09
RaceWRING0.88100.99262.2442e-08
RaceIsoAct0.80220.99326.1190e-08

Age. Normalization improves SPD’s debiasing, while IsoAct achieves the highest DD and retrieval retention among the compared methods.

Gender. IsoAct exceeds WRING on both metrics and achieves higher retrieval retention than SPD+Norm at a similar DD. Normalization reduces SPD’s DD from 0.8243 to 0.4289 despite restoring near-zero unit-norm error.

Race. SPD+Norm and WRING achieve higher DD, while IsoAct achieves higher mean retrieval retention.

These comparisons address both the normalization control and the closely related rotation-based method you highlighted. Since SPD+Norm, WRING, and IsoAct all achieve near-zero unit-norm error, these results strengthen the empirical case for IsoAct beyond satisfying the output constraint.

2. Isolating the contribution of rotational editing on KLIFS.

To address your observation that the joint SE(3) results demonstrate "feasibility rather than the independent benefit of rotational editing," we conducted a controlled ablation.

Starting from the joint configuration in our previous response, we restored the original rotation of each edited pose (RJ,tJ)(R_J,t_J) to obtain (R,tJ)(R,t_J). The two conditions differ only in rotation; translation outputs are exactly identical, and all other experimental settings are unchanged.

N.Acc and N.macro-F1 measure kinase-group prediction. Both conditions remain above Null, so lower nuisance scores are better here; higher task scores are better. Entries are full-test estimates ± SD from 1,000 paired test bootstrap resamples. Bold marks the better point estimate.

ConfigurationN.AccN.macro-F1aC F1aC RecallaC AUROCPocket F1Pocket RecallPocket AUROC
Original rotations restored0.1314 ± 0.00660.1302 ± 0.00660.3907 ± 0.02510.3391 ± 0.02550.7262 ± 0.01480.8355 ± 0.01150.8203 ± 0.01140.9717 ± 0.0021
Joint editing0.1283 ± 0.00670.1167 ± 0.00630.3922 ± 0.02480.3448 ± 0.02530.7382 ± 0.01410.8288 ± 0.01290.8241 ± 0.01220.9722 ± 0.0022

Keeping the edited rotations brings both nuisance scores closer to Null, with macro-F1 decreasing from 0.1302 to 0.1167. It also yields higher point estimates for all three aC metrics and Pocket Recall/AUROC; Pocket F1 is lower (0.8288 versus 0.8355).

Thus, rotational editing contributes to nuisance reduction and favorable downstream performance with translation held fixed. This isolates its additional contribution beyond the feasibility result in our previous response.

We sincerely appreciate the time and attention you have devoted to our paper throughout the discussion. Your feedback has helped us strengthen the empirical evaluation and clarify the scope and contributions of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.