acceptodds
Under review as a conference paper at ICLR 2027

Tilted Mondrian: Triage and Mitigation of Fine-grained Disparities in Conformal Prediction

Abstract

Fairness in conformal prediction is often pursued through Mondrian calibration, which provides finite-sample coverage guarantees separately across groups. These guarantees, however, are coarse: substantial disparities can remain within finer strata. This coarse-to-fine fairness issue raises two questions: how severe can such disparities be, and how can they be mitigated? Aggregate disparities can depend on how groups are distributed across strata. For severity assessment, we introduce propensity standardization to examine the effect of population composition, and show how group composition can induce disagreement between aggregate and finer-stratum comparisons. For mitigation, we develop Tilted Mondrian to improve fairness compared to group-Mondrian at a finer level without additional model training. Tilted Mondrian retains groupwise coverage guarantees and mitigates coverage and prediction set size disparities within finer strata. Finally, we derive a method-agnostic bound that quantifies how much reduction in finer-stratum coverage disparity can be achieved.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
conformal prediction, fairness, uncertainty quantification, Mondrian conformal prediction, prediction set size, propensity standardization, subgroup disparities

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PBu2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies fairness in conformal prediction, focusing on differences that may remain within smaller subgroups even when coverage is maintained for larger groups. The authors first propose a way to examine how differences in group composition affect prediction set sizes. They then introduce Tilted Mondrian, which adjusts prediction scores to reduce disparities within smaller subgroups while maintaining group-level coverage guarantees.

Critical strengths
  • The paper addresses an important limitation of group-level fairness in conformal prediction. Even when larger groups receive similar coverage, smaller subgroups may still experience substantial differences. I think this is a meaningful problem, and the paper provides a useful way to study it. I also appreciate the theoretical analysis of how much these differences can be reduced while limiting coverage loss.

  • I appreciate that the proposed method aims to reduce disparities without requiring the model to be trained again. The authors provide theoretical support for maintaining group-level coverage and evaluate the method across different settings and repeated data splits. The experiments show promising improvements in several cases, which supports the practical value of the approach.

Critical weaknesses
  • I understand why the paper uses information from the model's predictions for standardization, and the authors provide some explanation for this choice. However, these features are also closely related to prediction set size. This makes it difficult for me to understand how much of the reported reduction in disagreement comes from differences in subgroup composition and how much comes from comparing samples with similar prediction confidence. I think the paper would benefit from showing whether the results remain similar when different features are used for standardization.

  • I found that the presentation of the experiments could be substantially improved. The main paper presents the primary results, but many useful comparisons and additional experiments are placed in the appendix. I had to spend more time going through different sections of the appendix to understand how consistent the improvements are and how the method compares with other approaches. I think some of these results should be included in the main paper, especially the comparisons and experiments that directly support the main claims. The presentation of the results could also be improved to make it easier to see where the method performs well and where it does not.

Questions for authors

1- Please explain how much the choice of prediction features affects the standardization results? Since these features are closely related to prediction set size, I would like to understand whether the reported reductions remain similar when different features are used.

2- Please provide a more direct comparison with subgroup-level calibration using the same test coverage gap measures? I appreciate the existing comparison of coverage loss and prediction set sizes, but I think comparing the main fairness measure under similar coverage loss limits would help show the additional benefit of the proposed method.

3- Please provide a more explicit summary of how consistently the method improves subgroup fairness across different experiments and data splits? I found the main results promising, but many important results are in the appendix, making it difficult to understand the overall performance of the method and how much improvement it provides.

2 comments

We thank the reviewer for the thoughtful and constructive feedback and for recognizing the importance of the coarse-to-fine fairness problem, our theoretical analysis of mitigation limits, and the practical value of Tilted Mondrian (TM). We address the concerns below with new profile-sensitivity experiments, direct fairness–harm comparisons with Direct Cell Mondrian, and a summary of empirical evidence across datasets and resplits.

W1/Q1. Robustness to the Choice of Standardization Profile

The profile HH defines the standardized population QHQ_H, allowing us to assess how population composition influences aggregate–stratum fairness comparisons. Prediction confidence, when included in HH, represents one component of the composition being standardized.

New sensitivity experiments (see our response to Reviewer 6vX4, Q4). We evaluated three distinct profiles: confidence-only, predicted-label-only, and full profiles, holding the classifier, conformal thresholds, prediction sets, and evaluation splits fixed. See Section 6.1; Appendix D, Table 15. The results show that conflict-strength reductions persist across profiles constructed from different subsets of prediction-derived features. The choice of HH determines the standardized population comparison.

W2/Q3. Empirical Consistency and Presentation of Results

We thank the reviewer for this constructive suggestion. To make our empirical findings easier to assess, we summarize the key conclusions from the main paper, appendices, and additional experiments; see revised Section 6.2:

  • Coverage disparity mitigation (Table 2; Appendix F, Table 28). TM achieves positive mean test worst-gap reductions in 10/16 dataset–score configurations, with improvements in 230/320 resplits.

  • Consistency across datasets (Appendix E.1, Table 24; Appendix F, Table 28). Test worst-gap improvements occur in 71/80 resplits for Bias in Bios, 80/80 for Folktables occupation, 57/80 for Folktables income, and 22/80 for ATUS. Improvements are especially consistent in the first two datasets, while Folktables income exhibits score-dependent gains and ATUS presents smaller initial disparities.

  • Direct Cell Mondrian comparison (Appendix E.1, Tables 21–23; c4zs, W3). TM achieves substantially lower mean maximum baseline-relative coverage harm (0.82 vs. 5.55 pp) and lower set-size dispersion in 15/16 configurations.

  • Fairness–harm trade-off (Appendix E.1, Table 24). Under the same test criterion of positive worst-gap reduction and baseline-relative coverage harm ≤ 1 pp, TM satisfies both conditions in 143/320 resplits versus 18/320 for Direct Cell Mondrian.

  • Propensity standardization (Table 1; Appendix D, Table 15; Appendix F, Table 27; 6vX4, Q4). Conflict strength decreases in 899/900 evaluations using the original standardization profile across repeated data splits. The sensitivity analysis further characterizes how reversal counts vary with the choice of HH.

  • Additional fairness interventions (Table 3; Appendix E.2; Appendix G). Set-size-targeted TM reduces test worst set-size gaps in 234/320 resplits, with accompanying coverage–set-size trade-offs. Boosted TM further illustrates the flexibility of the framework through covariate-adaptive score interventions.

Q2. Fine-grained Fairness and Coverage-Harm Trade-offs

Direct Cell Mondrian calibrates explicitly within (G,Z)(G,Z) cells, whereas TM retains GG-level calibration and introduces finer-stratum interventions. Their raw coverage gaps therefore reflect different calibration objectives. We compare their results using the same test coverage-gap measure alongside baseline-relative coverage harm.

1. Direct comparison across datasets and scores.

We added direct comparisons in Section 6.2 (Table 4) and Appendix E.1 (Tables 23–24), illustrating their different fairness–harm trade-offs.

2. The additional benefit of Tilted Mondrian.

Our contribution is to separate the granularity of fairness intervention from that of conformal calibration. TM provides an additional finer-stratum intervention layer while retaining the original Group-Mondrian calibration and its finite-sample groupwise validity guarantee. The null intervention θ=0\theta=0 remains feasible on the tilt split, allowing the procedure to preserve the baseline when further adjustment offers limited benefit.

Thus, Direct Cell Mondrian strengthens cellwise coverage guarantees, while TM provides an approach that seeks to reduce finer-stratum disparities subject to empirical coverage-loss constraints. These comparisons make the practical value and trade-offs of the proposed intervention more explicit (see also our response to c4zs, W3).

Thanks for your thoughtful answers. Thanks for adding bullet-points to your empirical part in the paper. I will keep my assessment score and will consider your answers when discussing with other reviewers.

Best,

Reviewer_c4zsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes tilted Mondrian conformal prediction, which is a technique for adjusting conformal prediction in a group wise setting when there is imbalance within subgroups so that all subgroups receive more comparably-sized conformal sets. The authors discuss various metrics for understanding this imbalance in various strata beyond main disaggregation, including propensity reweigthing to a shared support distribution. Empirically, they demonstrate that this approach can improve a number of metrics around imbalance across datasets.

Critical strengths
  • I found the analysis of imbalance metrics in sections 3-4 useful and interesting
Critical weaknesses
  • hard to parse in points- I got fairly confused in the introduction and then around Sec 5 when the tilted Mondrian method is introduced (for instance, how are the shift values theta + and - selected?)
  • need more background on the original Mondrian method - hard to know otherwise what the novelty is
  • would be good to compare to the regular Mondrian baselines applied to the subgroups (as mentioned on L348), both conceptually and empirically
Questions for authors
  • what is the novelty of the method compared to original Mondrian? how does it compare empirically?
2 comments

We thank the reviewer for recognizing the value of our imbalance analysis. We clarify how Tilted Mondrian addresses finer-stratum disparities, highlight its methodological novelty, and provide a more direct comparison with standard Mondrian.

W1. Clarification of the shift-selection procedure

We have added clarifications in Section 5.1. These empirical constraints in Algorithm 2 bound coverage loss in initially advantaged cells, protect disadvantaged cells, and prevent gap increases. Among feasible candidates, we minimize the worst stratum gap, using the lexicographic tie-breaks specified in Algorithm 2 and Algorithm 4(Appendix E.1). Finally, the selected score is frozen before independent group-Mondrian calibration on Dcal\mathcal D_{\mathrm{cal}}, retaining finite-sample groupwise validity (Theorem 4).

W2 / Q1. Novelty relative to standard Mondrian CP

1. Background: Mondrian CP We have added a brief introduction to standard Mondrian CP in Section 5.1.

2. Motivation: Small and imbalanced calibration cells

Figure 4 of [A] illustrates how pooling heterogeneous datasets can obscure within-dataset associations between a substantive fairness metric and coverage/set-size gap. This broader concern that aggregate summaries can obscure finer-level patterns motivates our study of cross-group disparities within strata despite groupwise coverage guarantees. One approach to deal with finer-stratum disparity is Direct Cell Mondrian, which replaces q^g\hat q_g with q^g,z\hat q_{g,z}, providing coverage conditional on each supported (G,Z)(G,Z) cell. Its price is calibration set partitioned into small cells with sizes ng,zn_{g,z}.

Illustrative example. Suppose each gender has 1,000 calibration observations, but one occupation contains 20 women and 800 men. The finite-sample conformal coverage grid has spacing of approximately 1/(n+1)1/(n+1):

  • Women (20): 1/21≈4.761/21\approx4.76\\%.
  • Men (800): 1/801≈0.121/801\approx0.12\\%.
  • Coarse group (1,000): 1/1001≈0.101/1001\approx0.10\\%.

Small cells can yield coarser quantiles, greater variability, and larger prediction sets. This motivates finer-stratum disparity mitigation without requiring cellwise calibration.

3. Our methodological contribution

We identify finer-stratum disparities masked by aggregation and introduce Tilted Mondrian to mitigate them while preserving groupwise validity. Tilted Mondrian adjusts scores within finer cells while retaining group-level conformal calibration:

For a chosen coarse partition GG, the method targets disparities in finer (G,Z)(G,Z) cells while estimating each final threshold from the GG-level calibration pool ngn_g. Score selection and groupwise recalibration retain the original groupwise validity guarantee (Theorem 4). This is a general, post-hoc intervention at one finer level, without requiring a separate final calibration threshold for every cell. Compared to Mondrian CP, our method reduces finer-stratum disparity while retaining ordinary groupwise calibration, without requiring separate calibration thresholds for every finer cell.

W3. Comparison with direct cell Mondrian

The three procedures serve different calibration and intervention objectives. Here ngn_g and ng,zn_{g,z} refer to observations in the final calibration split. Tilted Mondrian also needs the separate tilt split. We add a conceptual comparison table in Appendix E.1 (Table 20). Empirical comparisons appear in Section 6.2 (Table 4) and Appendix E.1 (Tables 21–24).

Direct Cell Mondrian is the appropriate choice when a finite-sample guarantee within each cell is essential. We include Direct Cell Mondrian as a reference method, while recognizing that it targets a different coverage guarantee. More generally, Tilted Mondrian is not restricted to a particular Mondrian partition. It can be applied at any prespecified calibration level to target disparities across a finer refinement, while preserving the coverage guarantees associated with the original calibration partition.

Although Direct Cell Mondrian provides stronger cellwise coverage guarantees, partitioning the calibration data into imbalanced (G,Z)(G,Z) cells introduces unequal calibration precision. Smaller cells have coarser conformal quantile resolution and greater sampling variability, potentially leading to inflated prediction sets. Tilted Mondrian avoids this finer-cell fragmentation by retaining pooled groupwise calibration while enabling targeted disparity mitigation.

Tilted Mondrian is designed to enhance standard Group-Mondrian in settings where finer-stratum disparities persist. It preserves the original groupwise validity guarantee while offering an additional mechanism for improving fairness. Direct Cell Mondrian remains valuable when cellwise coverage guarantees are essential. Our approach provides a complementary option with a different fairness–efficiency trade-off.

[A] Liu et al. “Beyond Procedure: Substantive Fairness in Conformal Prediction” ICML 2026

Reviewer_c4zsOpenReview ↗

Response

Thanks for the rebuttal. My inclination is to keep my score as is but I will reconsider in conversation with other reviewers.

Reviewer_6vX4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Group-Mondrian CP equalizes coverage across groups GG but can leave disparities within finer strata ZZ. The paper offers (1) propensity standardization, an overlap-weighted re-averaging of delivered set sizes to test whether aggregate–stratum disagreements are composition effects, with prediction sets left unchanged; and (2) Tilted Mondrian, cell-specific constant score shifts selected on an independent tilt split under harm/gap constraints, followed by standard group-Mondrian recalibration. Theory: an overlap-weight balance property (Prop. 2), groupwise validity of the tilted score (Thm. 4), existence of a feasible tilt (Thm. 5), and a method-agnostic bound on per-stratum gap reduction (Thm. 6). Experiments on Bias in Bios, Folktables (occupation, income), and ATUS with SIMPLE/APS/RAPS/SAPS scores show reduced worst stratum coverage gaps on Bias in Bios and Folktables occupation, with mixed results elsewhere.

Critical strengths
  1. A guarantee-preserving mitigation layer that avoids cell-wise splitting. Selecting shifts on a tilt split, freezing, and recalibrating at the group level keeps the exact finite-sample groupwise guarantee while acting on finer cells, without retraining or partitioning calibration data into small cells. Against direct (G,Z)(G,Z)-cell calibration, Tilted Mondrian yields lower set-size dispersion in 15/16 configurations and far fewer >1pp cell-level coverage losses (103/320 vs 289/320 resplits). Where disparity exists it is materially reduced (Bias in Bios SIMPLE worst gap 0.161→0.0900.161 \to 0.090). This is the consideration that most supports acceptance.

  2. Candid, thorough empirical reporting. The appendix reports resplit variability, the active term of the Thm. 6 bound per configuration, ϵ\epsilon-sensitivity, and cross-metric side effects (coverage-targeted vs set-size-targeted mitigation) rather than only headline wins.

Critical weaknesses
  1. The theoretical contributions are largely inherited and do not cover the quantity the method targets. Proposition 2 is the exact-balance property of overlap weights (Li et al., 2018), restated. Theorem 4 holds for any score fixed on an independent split; the tilt contributes nothing to validity, and the paper's own take-away is that it merely "maintains" the group-Mondrian guarantee. Theorem 6 is a correct coverage-accounting identity, but in 11 of 16 dataset–score configurations (Table 16) the active term is the baseline gap dzd_z itself, i.e., the bound only says a gap cannot shrink by more than itself. There is no guarantee on cell-level gaps on the test distribution; feasibility is enforced in-sample on Dtilt\mathcal{D}_{\text{tilt}}, and the authors acknowledge test gaps can increase and that the approach is heuristic. For a paper whose stated contribution includes "mitigation limits," this gap between what is proved and what is claimed weighs heavily on my recommendation.

  2. Mitigation evidence is mixed and the cost is non-trivial. On the test split the mean worst-gap reduction is negative for Folktables income (SIMPLE, SAPS) and for all four ATUS scores (Table 23); robust gains are confined to Bias in Bios and Folktables occupation. The procedure halves the calibration pool to create Dtilt\mathcal{D}_{\text{tilt}}, precisely the regime (small strata) where finer-grained fairness is hardest. The paper cites conditional/multivalid calibration (Jung et al., 2022; Gibbs et al., 2025; Bairaktari et al., 2025) and score-transformation methods as related, but benchmarks only against cell-wise Mondrian and its own boosted variant, so it is unclear whether Tilted Mondrian is preferable to approaches that deliver (approximate) cell-level validity directly.

Questions for authors
  1. Theorem 4 holds for any score fixed on an independent split. What, if anything, does Tilted Mondrian guarantee about cell-level coverage gaps on the test distribution, e.g., a high-probability transfer of the Dtilt\mathcal{D}_{\text{tilt}} feasibility constraints to test as a function of cell sizes, even asymptotically?
  2. In 11/16 configurations the Theorem 6 bound reduces to dzd_z (Table 16). In what regime is the bound informative, and is it attainable (Theorem 5 only shows a nonzero feasible tilt exists)?
  3. Why were conditional/multivalid calibration methods (Gibbs et al., 2025; Jung et al., 2022) not included as baselines, given they target the same finer-stratum coverage and are cited as related?
  4. The standardization conclusions rely on ηH\eta_H being small, yet residual standardized cell total variation is 0.045 (Folktables occupation) and 0.091 (ATUS). How sensitive are the reversal / conflict-strength reductions to the choice of profile HH (4 features for ATUS vs 37 elsewhere)?
  5. How does mitigation degrade as ∣Dtilt∣|\mathcal{D}_{\text{tilt}}| or the smallest cell shrinks? Table 5 uses tilt splits of roughly 44k–102k observations.
1 comment

We thank the reviewer for the careful assessment and for recognizing the value of our guarantee-preserving mitigation framework and thorough empirical reporting. We particularly appreciate the observation that the comparison with Direct Cell Mondrian is a strong argument in favor of the paper.

W1 / Q1. Theoretical Guarantees on Test Set

Li et al. (2018, Theorem 3) establish finite-sample mean balance under logistic-regression MLE; our Proposition 2 provides a more general population-level balance identity for any integrable function of HH. We agree that Theorem 4 builds on classical group-Mondrian validity, but its purpose is to establish that the proposed, data-adaptive score adjustment preserves this guarantee when learned on an independent tilt split and frozen before calibration. It therefore provides the validity foundation for Tilted Mondrian. Our new DKW analysis provides a high-probability lower bound on population disparity reduction (Section 5.2, Proposition 7).

W1 / Q2. Informativeness and Attainability of Theorem 6

Theorem 6 is a method-agnostic bound on achievable gap reduction under coverage-loss constraints, not a claim of universal improvement. Its ϵ\epsilon-budget term is active in 5/16 configurations (Appendix E.1, Table 18). For Folktables occupation (APS), it is 0.0395 versus the baseline-gap ceiling of 0.0963; Tilted Mondrian achieves 0.0211 on the tilt split (Table 19). Attainability example. Consider two equally weighted strata with baseline coverages ca0=(0.99,0.81)c_a^0=(0.99,0.81) and cb0=(0.81,0.99)c_b^0=(0.81,0.99). With ϵ=0.01\epsilon=0.01, the feasible redistribution caθ=(0.98,0.82)c_a^\theta=(0.98,0.82) and cbθ=(0.82,0.98)c_b^\theta=(0.82,0.98) preserves each group's coverage at 0.90, respects cellwise harm constraints, and reduces both gaps by 0.02=2ϵ0.02=2\epsilon, attaining the bound. Suitable continuous score distributions admit such directional shifts. This verifies attainability in the regime.

W2 / Q3. Multivalid CP

We appreciate the contributions of [A, B] to conditional and multivalid conformal prediction. Our objective is to reduce cross-group disparities within strata relative to a group-Mondrian baseline, subject to explicit limits on coverage deterioration, while retaining groupwise final calibration. Conditional/multivalid coverage guarantees do not, by themselves, impose these baseline-relative constraints. For example, suppose that within one occupation, baseline coverage is 0.99 for women and 0.81 for men (nominal coverage 0.90). Bringing both to 0.90 would eliminate the disparity but reduce women’s existing coverage by nine percentage points. Our ϵ=0.01\epsilon=0.01 constraint permits at most a one-percentage-point loss for the initially higher-coverage cell and no loss for the initially lower-coverage cell, as evaluated on the tilt split. Final calibration remains groupwise. Thus, conditional validity and harm-constrained disparity mitigation are related but distinct fairness objectives. For a Direct Cell Mondrian comparison, see Appendix E.1 (Tables 20, 23–24) and our response to c4zs (W3).

Q4 / Q5. Sensitivity Analysis

Across three profiles (confidence-only, predicted-label-only, and full; Section 6.1 and Appendix D, Table 15), standardization reduces conflict strength in 54/54 evaluations; reversal counts increase in 11/54. The DKW bound in W1/Q1 weakens for smaller cells, whereas exact groupwise validity remains (Theorem 4). Empirical cell coverage changes in increments of 1/ng,ztilt1/n_{g,z}^{\mathrm{tilt}}: with ϵ=0.01\epsilon=0.01, fewer than 100 observations cannot accommodate even a one-observation coverage loss under the harm constraint. The null tilt is possible (Theorem 6). Resplit variability is reported in Appendix F (Table 28).

W2. Results and Data Cost

TM improves mean test worst gap in the 10/16 dataset–score settings and in 230/320 resplits (Appendix F, Table 28). Both Folktables income APS/RAPS improve in all 20 resplits. All six negative-mean settings have small baseline gaps (0.019–0.026), versus 0.041–0.179 for the ten positive-mean settings. Limited room for improvement can be obscured by tilted estimation and calibration noise. The new fairness–harm analysis is detailed in Section 6.2 (Table 4) and Appendix E.1 (Tables 23–24); see PBu2 (Q2). We reserve about half of the original calibration for tilt split. Splitting preserves finite-sample groupwise validity by Theorem 4. Crucially, TM still calibrates one threshold per group instead of per finer cell, avoiding calibration fragmentation. Despite the split, Bias in Bios (SIMPLE) reduces the mean worst test gap from 0.172 to 0.104 in all 20 resplits. All four Folktables occupation scores improve in every resplit. Thus, finer-grained mitigation is feasible without requiring finer-cell calibration.

[A] Jung et al. “Batch multivalid conformal prediction” ICLR 2023

[B] Gibbs et al. “Conformal prediction with conditional guarantees” Journal of the Royal Statistical Society Series B: Statistical Methodology 2025

Comments on the paper

1 comment

We sincerely thank the reviewers for their constructive feedback. We appreciate their recognition of the paper's contributions and their thoughtful suggestions for improvement. We begin with a brief overview of how the paper’s motivation, methodology, and theoretical and empirical findings fit together. A revised manuscript has been updated with changes highlighted in blue.

  • Motivation. Liu et al. [A] show that pooling datasets can obscure within-dataset relationships between substantive fairness and coverage/set-size gaps. Analogously, Mondrian CP guarantees coverage at a chosen grouping level but can leave disparities within finer strata. Direct calibration in finer cells offers cellwise coverage but may fragment calibration samples into small, imbalanced cells. This motivates assessing such disagreement and reducing finer-stratum gaps while retaining the efficiency benefits of pooled calibration.

  • Assessment. Set-size curves and a gap decomposition reveal Simpson-type reversals obscured by aggregation. Propensity standardization reweights observations to assess composition effects. Theorem 3 analyzes improved aggregate–stratum alignment.

  • Solution. Tilted Mondrian extends Mondrian CP at any chosen grouping level, allowing practitioners to target disparities one level finer while retaining the original level's finite-sample coverage guarantee. For example, starting from GG-Mondrian, it can address disparities within (G,Z)(G,Z) while preserving GG-conditional coverage. Final calibration remains at the original grouping level, avoiding further fragmentation into small calibration cells. This retains the benefits of pooled calibration, preserving stable quantile estimation.

  • Theory and experiments. Theorems 4–6 address the validity of Tilted Mondrian, feasibility, and limits. We added a high-probability DKW bound about finer-stratum mitigation from Tilted Mondrian at the end of Section 5 in the revised manuscript. Extensive experiments cover four tasks, multiple scores, repeated splits, method comparisons, sensitivity analyses, and limitations.

[A] Liu et al. “Beyond Procedure: Substantive Fairness in Conformal Prediction” ICML 2026

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.