acceptodds
Under review as a conference paper at ICLR 2027

How reaching a Reliable Ethnicity and Gender Inference (REGI)?

Abstract

Fairness evaluation in face analysis systems (FAS) uses automatic demographic attribute inference (DAI), which in turn relies on predefined demographic segmentation. However, the validity of fairness auditing hinges on the reliability of the DAI process. We revisit this process by proposing a fully reproducible ethnicity+gender prediction pipeline (REGI) that includes curation of diversified training and test datasets and a transfer-learning-inspired model training strategy. We ground our approach in a per-identity bound on the probability that an image is assigned a different segment than its identity's mean embedding. We show that the decision class this bounds covers, log-likelihood ratios of von Mises-Fisher mixtures, is realized by an RBF-SVM head on a pretrained face recognition encoder. Its margins being driven by the coverage of the training set motivates further our diversified training set WEG. We audit this pipeline across three dimensions: accuracy, fairness, and a newly introduced notion of robustness, defined via intra-identity consistency. We compare our approach with seven competitive baselines across multiple datasets and training setups, including pipelines trained specifically for attribute prediction and generalist VLMs. Our pipeline outperforms all baselines for ethnicity, the more challenging attribute, across the three audited dimensions. REGI also ensures an interesting performance-efficiency balance, beating all the dedicated attribute predictors even with smaller computational budgets. Importantly, the newly proposed attribute training dataset generalizes better than existing counterparts, while our new test dataset, more challenging, is composed more suitably for evaluating the three audited dimensions. To promote reproducibility and adoption, we release the dataset metadata, full codebase, pretrained models, and evaluation toolkit.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
demographic attribute inference, spherical representation learning, algorithmic fairness, face fairness auditing, face recognition, transfer learning, annotator robustness evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Zo1eOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work seemingly aims to achieve reliable race and gender inference through a transfer-learning-inspired model training strategy and the constructed demographic attribute inference datasets. This work also claims that the proposed REGI simultaneously improves accuracy, fairness, and robustness compared with widely used systems.

Critical strengths
  1. This work seems to contain a lot of information but is not clear. Please correct me if I misunderstand.
Critical weaknesses
  1. I suspect this work was generated by an LLM, but the author didn't organize it properly. Garbled color blocks always appear at the top of Figure 1, and the main text is quite strange to read. This is detrimental to the sustainable development of academia. However, please correct me if my judgment is incorrect.
Questions for authors

If possible, could the author please explain this sentence in "Statement on the Use of Large Language Models": "LLMs were also used for editing assistance on the manuscript; all technical content, proofs, experiments, and claims are the authors'." This would help clarify any misunderstandings.

4 comments

On the structure of the paper We agree that the paper contains a lot of information. We believe that doing so is required when tackling such a sensitive matter, due to the very nature of the output of the annotators on one hand, and their wide use in the fairness literature for mitigating and auditing bias on the other hand. We are, however, surprised by the statement about our work organisation. It follows a standard structure in the literature: theoretical characterisation of the problematics tackled by the paper in Section 3, implementation of a design answering these problematics in Section 4, and experimental results with relevant ablatives in Section 5.

Figure 1. The colored blocks are a rendering bug of Chromium's built-in PDF viewer, not part of the figure (see the corresponding Chromium fix: https://chromium.googlesource.com/chromium/src/+/0942c4d6e2b66644ff5eb69e925dec86162bf60d). The figure displays correctly in any other PDF viewer. This bug is actually expected for figures exported from draw.io, i.e., it is, if anything, proof that the figure is human-made. To judge an article based on it does not look aligned with ICLR Guidelines.

Overall, we therefore remind the reviewer that the ICLR 2027 Reviewer Guidelines (https://iclr.cc/Conferences/2027/ReviewerGuidelines) ask reviewers to "read the paper carefully, check arguments in sufficient detail to convince yourself of the paper's main claims", and to "only discuss points that are actually likely to influence your final accept/reject decision", adding that even "pointing out typos [...] just add[s] to everyone's workload without changing the final conference program". A rendering artefact that depends on the PDF viewer is not even a typo, and it says nothing about the question the guidelines ask reviewers to answer: "Does the paper support the claims?" Beyond the rendering artefact, the Paper Strengths reported by the reviewer raise serious doubts about the seriousness with which they actually read the paper.

Use of LLMs. We use LLMs for editing because we are not native English speakers: they correct our grammar, spelling, and structural mistakes. The (now corrected) grammar error in the very title is a very good illustration of why we need them to correct us. All technical content, proofs, experiments, and claims are ours, as the statement says; we made it more explicit in the revision (l. 540–541).

Contributions. We summarise here the list of our contributions:

  • We define new taxonomy-agnostic robustness metrics, which we consider necessary (but not sufficient) for auditing any annotator.
  • We show that widely used annotators do not perform well along the three evaluation axes we deem important: accuracy, fairness, and our robustness metrics.
  • We provide a class of functions that comes with a robustness guarantee, and show that an SVM with an RBF kernel falls into this class.
  • We show that this guarantee can be expressed as an upper bound driven by a signal-to-noise ratio: the margin of an identity along the decision direction, divided by the spread of its images in that direction. When identities are modelled as isotropic vMFs, the bound decreases exponentially with the concentration of the images of the same person. Face-recognition angular-margin losses are expressly designed to shrink this spread and model identities in this way; we made this explicit in the revision (p. 4, l. 204–205).
  • We therefore present a face-recognition backbone with an SVM head as a theoretically motivated annotator.
  • We note that the margin, i.e., the numerator of this ratio, is theoretically directly linked to the diversity of the training data.
  • We thus create an annotator training set, WEG$^{tr},asubsetofWebFacediversifiednotonlywithrespecttotheoutputclasses(, a subset of WebFace diversified not only with respect to the output classes (g$), but also with respect to finer ethnicities (ee).
  • To do so, we use our own theoretically motivated sampler, which balances p(g)p(g) and p(e)p(e) simultaneously and preserves p(g∣e)p(g|e), which we consider a robust descriptor of the relation between ee and gg.
  • At the same time, we create WEG$^t$, a more challenging benchmark, diversified along gg and ee.
  • We empirically verify the impact of WEG$^{tr}$ and of the RBF-SVM head on accuracy, and show that both increase performance, which rules out the possibility that the face-recognition backbone alone drives the performance.
  • We compare the resulting annotator with existing ones: it reaches state-of-the-art performance for ethnicity inference and competitive performance for gender inference (second only to the much larger QWEN3 models), at a limited computational cost.
  • We show that the performance remains remarkably stable when changing the size and the type of the backbone, the size of the SVM training set, the size of the backbone training set, the balancing of the backbone training set, and the loss used to train the backbone.
  • Illustrating the importance of taxonomy-agnostic metrics, we also use our robustness metrics to compare REGI with existing models based on the skin-color taxonomy, which highlights the instability of these approaches.

We could also add the impact of the annotator performance on the estimation of the fairness of face-recognition models: low-performance annotation leads to an underestimation of the bias of the audited models. However, as highlighted by the reviewer, this work already "seems to contain a lot of information".

Reviewer_Zo1eOpenReview ↗

Official Comment by Reviewer Zo1e

Dear Authors,

Thank you for your further responses; they really help me and readers to further understand your work.

However, I still believe that this work has problems with its language and formatting for human readers, and these problems remain unresolved. These include, but are not limited to, some results in Figure 1 being obscured, and the font sizes in the appendix figures and tables varying significantly. This demonstrates the author's weak seriousness towards their work.

These are fundamental issues concerning the high quality publication standards. I believe this work will be good and meaningful if the relevant content is carefully revised, even if it is not currently done.

Best,
Reviewer Zo1e

Dear Reviewer, We take note of your comments and regret that they are not based more on specific elements of the article. We will leave it to the AC to determine whether your comments comply with the reviewer guidelines, as noted in our initial answer.

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

The paper proposes REGI, a pipeline for predicting four coarse perceived-ethnicity categories and two gender categories for demographic auditing. REGI combines a frozen face-recognition encoder with an RBF-SVM classifier. The paper also introduces WEG, a dataset curated from WebFace using identity-level label aggregation and sampling intended to improve fine-grained demographic coverage, and homogeneity entropy (HomE), which measures prediction consistency across images of the same person. The theoretical contribution comprises a distribution-free bound on identity-level prediction changes and an interpretation of RBF kernels through von Mises–Fisher mixtures. Experiments compare seven baselines on BFW, RFW, and WEG, with extensive ablations of training data, classification heads, backbones, and training strategies. The strongest results are on ethnicity: REGI-IR50 reports accuracies of 95.9%, 95.8%, and 90.8% on the three datasets, respectively, along with improved group-accuracy balance and consistency. Gender results are competitive, although Qwen models achieve higher accuracy and consistency.

Critical strengths
  1. Strong empirical results supported by broad ablations. The gains extend to existing benchmarks, not just the proposed WEG test set. Table 2 separates the effects of training data and the classification head, while Table 3 compares pretraining scale and training strategy. For example, with BUPT training data, replacing the fully connected head with an RBF-SVM improves ethnicity accuracy from 81.7% to 87.9% on WEG and from 90.5% to 95.6% on BFW. These comparisons make the practical contribution more convincing than comparing complete pipelines alone.
  2. A useful evaluation emphasizes identity-level consistency. HomE exposes prediction instability that aggregate accuracy can conceal. The paper evaluates this alongside accuracy and group disparities and explicitly acknowledges that consistent predictions can still be wrong (Appendix A.5). This matters for auditing workflows that rely on automatically assigned demographic labels.
Critical weaknesses
  1. The theoretical justification contains an incorrect claim and an unresolved connection to the implemented model. The necessity direction of Theorem 2 (p. 23) is false as stated. For example, let there be two groups and two subgroups, with p(g1∣e)=0.8p(g_1\mid e)=0.8 and p(g2∣e)=0.2p(g_2\mid e)=0.2 for both subgroups. Then S1=1.6S_1=1.6 and S2=0.4S_2=0.4, but the prescribed sampling still gives p(e∣g)=1/2p(e\mid g)=1/2 and hence uniform subgroup marginals. Thus equal group masses are sufficient but not necessary. Separately, we establish the RBF-SVM equivalence to the specified log-mixture decision rule for zero bias. With nonzero bias, the decision compares a difference of densities against an additive threshold, which is not generally a constant shift of their log ratio. The multiclass aggregation also needs an explicit argument. Appendix C evaluates a separately constructed mixture classifier rather than the trained SVM, so it does not resolve this gap. The generic Cantelli bound remains useful, but these issues weaken the claimed theoretical explanation for REGI.
  2. The evidence for reliable demographic ground truth remains insufficiently documented. WEG labels come from FFNet predictions, and their inclusion depends partly on within-identity agreement. The binomial argument assumes independent errors with a common accuracy; variation in pose or illumination does not establish that assumption, since systematic mistakes can persist across every image of a person. Moreover, Equation 28 describes the probability of a correct-class supermajority, rather than directly calculating correctness conditional on retention. The reported 95.5% aggregate-label accuracy on BUPT is materially below the independence-based figure exceeding 99.6%. Appendix D.1 importantly reports manual verification of WEG’s test labels and removal of fewer than 0.3% of samples. However, it does not adequately describe annotators, independent labeling, disagreement handling, or subgroup error rates. A removal rate alone does not establish residual label accuracy. The BFW and RFW results partly mitigate this concern, but the evidence does not yet support the stronger claim that aggregation resolves annotation circularity.
Questions for authors
  1. Clarification of implementation and theory: Does the deployed RBF-SVM use a nonzero intercept, and how are binary decisions combined for multiclass prediction? Please establish the claimed connection for that exact implementation and clarify whether Appendix C’s diagnostics hold for the trained REGI classifier.

  2. Correction of the sampling result: How should Theorem 2 be revised in light of the counterexample above? Does weakening necessity to sufficiency change any claimed justification or implementation of WEG sampling?

  3. Clarification of label validation: What was the manual verification protocol for WEG’s test set—coverage, annotator count, access to automatic labels, agreement, adjudication, and treatment of ambiguous cases? What evidence supports residual label accuracy by demographic group?

  4. Clarification of evaluation independence: Appendix H.5 describes “best RFW validation accuracy at optimal epoch.” Was RFW used to choose checkpoints or hyperparameters for any reported test results? If so, which results remain valid under selection exclusively on a separate validation set?

2 comments

Question 1 (deployed classifier and theory)

We thank the reviewer for the detailed review. To keep a dense derivation readable, the submitted version stated the theory for a zero intercept and an argmax rule, with the intercept presented as a calibration of the threshold. Appendix C therefore checked an explicit member of the class. We agree that the membership of the deployed head was stated more strongly than this simplified derivation supports. The revision now establishes the connection for the exact implementation.

Implementation. The deployed head is scikit-learn's SVC with an RBF kernel, γ=1/(D Var(X))\gamma=1/(D\,\mathrm{Var}(X)) and C=1C=1. It fits one intercept per pair of classes, predicts ethnicity by one-vs-one voting, and predicts gender with a single binary head (as now clarified in App. F, l. 1681–1692).

Intercept. On the sphere, a constant is a vMF mixture with uniformly distributed centers. The intercept therefore adds a uniform component to the mixture of the class it favors, and the head remains exactly a log-likelihood-ratio test between two vMF mixtures (cf. Proposition 2, added in App. B.5). Definition 1 is now generalized accordingly (l. 171–174), and Lemma 1 and Proposition 1 are unchanged.

Multiclass. The guarantee follows directly from applying our bound to each one-vs-one head. A flip requires the predicted class to lose at least one of its duels, so the flip probability is at most the sum of the binary bounds of these duels (see Corollary 1, added in App. B.6).

Appendix C on the trained classifier. We report, in the revised version, the full protocol on the REGI-IR50 head of Table 1 itself: its backbone and its one-vs-one SVM, intercepts included (Figure 4). Out of sample, the bound holds for 99.1% of the identities (95.1% for the min-class variant). The SNR ranks the flip rates with a Spearman correlation close to the reliability ceiling (l. 1169–1180 and l. 1191–1200). Only 27 of the 48,000 images lose a duel without flipping, so the vote adds almost no slack to the bound. The explicit rules are kept as stress tests.

The theoretical explanation therefore applies to the classifier we deploy. We thank the reviewer for noting that the Cantelli bound remains useful.

Question 2 (Theorem 2), Weakness 2 and Question 3 (labels)

Theorem 2. We apologise for the confusion. The statement in the submitted version could have made unclear a requirement that the proof used implicitly: we want the sampled data to also keep the conditional p(g∣e)p(g|e) intact, in order to reflect the dynamic between ee and gg. We revised the formulation of the theorem so that this condition appears clearly (cf. App. D.2 of the revised version). Keeping the conditional intact is the motivation for using the IPF constraint: it is the condition for the sampler to reproduce p∗(g∣e)p^*(g|e) with uniform fine-grained marginals. We also corrected the sentence of Sec. 4.1 accordingly (l. 297–299).

Label quality (Weakness 2). We revised Sec. 4.1 (l. 276–285) and App. D.1 (l. 1353–1366) accordingly to enhance clarity:

  • the binomial computation is now explicitly framed as an idealized bound that assumes independent errors.
  • Eq. (28) is distinguished from the accuracy of the retained labels, which the 95.5% on BUPT measures directly.
  • The remaining gap is attributed to different causes: correlated error, BUPT metadata own error, and taxonomy misalignment.

However, we do consider aggregation a strong way to mitigate circularity. On BUPT, our protocol reduces the error rate by almost 75%. The robustness of the labels is further confirmed by the results of REGI on BFW and RFW, whose labels are independent of FFNet: REGI beats all baselines and achieves state-of-the-art results on these benchmarks, even with a much smaller backbone (RN18) than the dedicated annotator baselines. (c.f. l. 1362-1366)

Q3, manual verification. We added precision about how the human verification was performed. Its role is only that of a sanity check, in which a single annotator inspected the automatic labels of 800 WEG$^t$ identities. We do not use this check as evidence of label accuracy, and App. D.1 now clearly describes it as such (l. 1367–1368).

Our confidence in the labels rests on the three elements above, and the results on BFW and RFW are the strongest of them. The RBF-SVM head fits the WEG$^{tr}$ labels almost perfectly (training accuracy close to 100%), so it would reproduce their systematic errors. Yet, on BFW and RFW, whose labels do not come from FFNet, REGI nearly halves the ethnicity error of FFNet with RN18 (from 12.1% to 6.0% on BFW, and from 11.9% to 6.6% on RFW), and divides it by almost three with IR50 (Table 1). This is hard to reconcile with training labels that inherit significant circularity with FFNet.

Question 4 (evaluation independence) and ethics

Q4. RFW was never used to select a checkpoint or a hyperparameter. For each backbone, the checkpoint is the epoch with the best accuracy of the REGI head on the validation split WEG$^v$. WEG$^v$ is identity-disjoint from the test sets, which are evaluated only at that epoch (we made this clearer in App. F, l. 1687–1691). The sole purpose of H.5 is to be an a posteriori analysis that selects no model (now stated explicitly, l. 2097–2099). The objective of H.5 is thus to audit REGI, to help the community understand what impacts our proposed framework and to what extent.

Ethics. We acknowledge the concerns raised by inferring demographic attributes from faces, and we take them seriously. We nevertheless consider tools like REGI essential to any fairness audit and bias mitigation approach. Auditing a face-recognition system for instance means measuring its performance per demographic group, so the audit is only as reliable as the demographic labels it relies on. Turning a blind eye to the limited performance of existing annotators would therefore lead to an underestimation of the biases in the audited systems. In face recognition, these biases have documented and unacceptable consequences, such as the denial of services or rights and the wrongful arrest of innocent people.

  • Provenance and consent. WEG is a resampling of WebFace42M and inherits its terms. We release only identity lists and metadata, never images.
  • Identity-linked inferred attributes. The metadata, namely the inferred gender and coarse ethnicity, are needed to train and audit annotators. They are algorithmic estimates, not statements about any individual's identity.
  • Safeguards. Code, models, and metadata are released under a restricted-use license. It prohibits demographic attribute inference as a primary objective, standalone task, or service (e.g., targeted marketing or profiling), and permits it only to quantify or mitigate bias. The Ethics statement now says so (p. 10, l. 509–513).

The sole purpose of this work is, again, to make bias audits and mitigation more reliable: underestimating the bias of face analysis systems has had serious real-world consequences, which is often one of the main motivations for fairness-related works.

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

This paper proposes REGI, a demographic attribute inference pipeline combining a pretrained face recognition encoder with an RBF-SVM classifier. It introduces the WEG dataset, a theoretical analysis of intra-identity prediction consistency, and homogeneity entropy (HomE) as a robustness metric. Experiments on multiple datasets show promising improvements in ethnicity inference accuracy, fairness, and robustness, with a favorable performance–efficiency trade-off. 1.The paper highlights the importance of reliable demographic attribute inference for valid fairness auditing and introduces intra-identity consistency as a complementary evaluation dimension.

  1. The combination of a pretrained face recognition encoder and an RBF-SVM is theoretically motivated, while the proposed WEG dataset emphasizes demographic diversity and fine-grained subgroup coverage.

  2. Experiments across multiple baselines, datasets, architectures, and training configurations demonstrate promising performance improvements and a favorable computational efficiency trade-off, particularly for ethnicity inference.

Critical strengths

None

Critical weaknesses

1.The connection between the RBF-SVM and the proposed vMF-mixture decision class remains unclear for nonzero bias terms and the actual multiclass SVC implementation. This raises concerns about whether the theoretical analysis applies to the proposed REGI classifier 2.The paper contains several inconsistencies in table captions, reported computational costs, metric definitions, and scaling conventions. These issues reduce the clarity and suggest that the manuscript requires more careful verification.

Questions for authors
  1. In Formula 5 and Formula 25. The theoretical connection between the RBF-SVM and vMF mixture models requires further clarification. While the RBF kernel on normalized embeddings is proportional to a vMF density, the subsequent equivalence appears to rely on a zero bias term. For a nonzero SVM bias, the decision rule involves a threshold on the difference between mixture densities, which is not generally equivalent to a fixed-threshold likelihood-ratio test.

  2. Definition 1 (p. 4) formulates multiclass prediction as an argmax over class-specific log-sum-exp scores, whereas the standard scikit-learn SVC uses a one-vs-one strategy for multiclass classification. These decision rules are not generally equivalent. Please clarify how the theoretical analysis applies to the actual REGI classifier.

  3. Macro-averaging (MaMA) typically involves calculating a metric for each statistical unit and then averaging these results with equal weights, whereas micro-averaging (MiMA) aggregates the raw counts before computing the overall metric. In Appendix G.3, Equations (38) and (39) appear to implement formulations that are the reverse of their stated labels. The authors are requested to re-verify these definitions and ensure that the formulas are consistent with them.

  4. Table 1 reports 8767 GFLOPs for QWEN3-8B, whereas in Table 5, 8767 corresponds to the number of parameters (M), and the reported computational cost is 1973 GFLOPs. Please verify and correct the value in Table 1.

  5. The caption states that “Gender columns are left blank,” but the table actually contains values for gender inference accuracy, fairness, and robustness. Please revise the caption accordingly.

  6. The caption states that the comparison is conducted using an IR50 backbone, but the table includes multiple architectures, such as IR18, IR50, RN101, RN34, RN50, ViT-S, and ViT-T. Please revise the caption to accurately reflect the contents of the table. In addition, please check whether “Imbb” under WEGt is a typo for “Imb.”

  7. Ensure that metric definitions and scaling conventions are consistent across Table 1, Table 4, and other relevant tables to avoid ambiguity in interpreting the reported values. For example, the fairness metric should be clearly labeled as “Fairness (DoB × 100)” wherever this scaling is applied.

1 comment

Theory and the deployed classifier (Weakness 1, Formulas 5 and 25, Definition 1)

We thank the reviewer for these precise questions. To keep a dense derivation readable, the submitted version deliberately stated the theory for a zero intercept and an argmax rule, and presented the intercept as a calibration of the decision threshold. We agree that the sentence stating that every binary head belongs to Fκb\mathcal F_{\kappa_b} was stronger than this simplified derivation supports. The revision resolves both points exactly, without changing the theory.

Nonzero intercept. The reviewer is right: for b≠0b\neq0, the head compares the two support-vector densities against an additive threshold. On the sphere, however, a constant is itself a vMF mixture, namely one whose centers are uniformly distributed (l. 1026–1036). The intercept therefore adds a uniform component to the mixture of the class it favors. The head is then exactly a log-likelihood-ratio test between two vMF mixtures (Proposition 2, App. B.5, l. 1037–1065). Definition 1 now allows general mixing distributions (l. 171–174), and Lemma 1 and Proposition 1 hold unchanged.

One-vs-one voting. The reviewer is correct that the deployed head votes over one-vs-one binary heads instead of taking the argmax of Definition 1. The guarantee follows directly from applying our bound to each of these binary heads. A flip requires the predicted class to lose at least one of its duels, so the flip probability is at most the sum of the binary bounds of these duels (Corollary 1, App. B.6, l. 1074–1107). App. F now describes the deployed head exactly (l. 1682–1686).

Presentation (Weakness 2, Questions 3 to 7)

Thank you for the careful check. We corrected every point and re-checked the other tables and captions.

  • Q3, MaMA/MiMA. Eqs. (38) and (39) implement our verbal definitions. In our naming, "macro" and "micro" refer to the scale of aggregation (dataset vs. identity). We now state this explicitly (App. G.3, l. 1836 and l. 1850–1855). If the reviewer prefers, we will swap the names in the final version.
  • Q4, QWEN3-8B. This was a copy error. We corrected the cost to 1973 GFLOPs in Table 1 and in Table 6, using the correct value as reported in Table 5.
  • Q5, Table 4. We removed the internal-review artifact left in the caption.
  • Q6, Table 9. The caption now describes the seven backbones, "Imbb" now reads "Imb", and the row names match Table 7.
  • Q7, scaling. All tables indeed use the same convention. It is now stated in App. G.1 and G.2 for both performance and fairness metrics.

Comments on the paper

1 comment

We thank the reviewers for their valuable comments. The revised PDF shows every change in blue, and our replies refer to the associated line numbers (with l. PageNumber). The main changes are:

  1. Theory and deployed classifier (PFWH, enAt). The RBF-SVM head with a nonzero intercept belongs exactly to the decision class (new Proposition 2, App. B.5). The flip bound holds for scikit-learn's one-vs-one vote (new Corollary 1, App. B.6). No other result changes.
  2. Validation on the trained head (enAt). Appendix C now includes the REGI-IR50 head of Table 1 itself.
  3. Theorem 2 (enAt). The theorem is restated with its unclear second hypothesis, hence justifying necessity and sufficiency.
  4. Label quality (enAt). The binomial argument is indeed an idealized bound. We try to go beyond this idealization by reporting the accuracy of our protocol on BUPT, and furthermore, by testing REGI on unrelated benchmarks (RFW and BFW).
  5. Presentation (PFWH). Minor Corrections (i.e. GFLOPs of QWEN3-8B, captions of Tables 4 and 9, metric scaling (App. G), MaMA/MiMA convention).
  6. Evaluation protocol and ethics (enAt). Checkpoint selection is performed on WEG$^v$ only (App. F), we change the wording of App. H.5, and clarified our stance on released metadata and license (Ethics statement).
  7. LLM statement, structure and contributions (Zo1e). We stated exactly what LLMs were used for, described our paper's structure (which we think is generic in the literature), and made a summary of our contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.