acceptodds
Under review as a conference paper at ICLR 2027

Lebesgue-Contrastive Prediction for Multivariate Conformal Regression

Abstract

The minimum-volume prediction region with conditional coverage is a superlevel set of the conditional density, known as the conditional highest-density region (HDR), so a scalar score reproducing this level-set geometry suffices for optimal multivariate regression. We propose Lebesgue-contrastive prediction (LECP), which learns such a score by logistic discrimination of observed responses against a Lebesgue reference measure, requiring neither a normalized density model nor an invertible parameterization. An alternating procedure fits the contrastive score and a conditional-quantile level that tracks the target density contour, followed by split conformal calibration on an independent hold-out set. The resulting regions enjoy finite-sample marginal coverage by construction and, under estimation consistency, converge to the conditional HDR. On synthetic and real multivariate regression benchmarks, LECP yields substantial volume reductions over existing methods while maintaining competitive conditional reliability.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
conformal prediction, multivariate regression, minimum-volume sets, density level sets, contrastive 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_tJZ8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The conditional minimum volume problem (minimizing Lebesgue measure for lower-bounded coverage), has a known solution: the highest density region. The key idea of this paper is to realize that the conditional density can only be estimated up to a constant (which is enough to preserve the level-set geometry), if a proper correction to obtain correct quantiles is performed. Then classical split conformal prediction is used. The unnormalized conditional density estimation is performed using an adaptation of a contrastive loss, while the traditional pinball loss is used for quantile regression. Theoretical results based on existing statistical methodology accompany the results, as well as experiments validating the approach.

Critical strengths
  • The idea of unnormalized density estimation is new and elegant and allows to use stable existing tools for such density estimation (the other step of correcting a score using pinball loss was already present in Plassier et al 2025).
Critical weaknesses
  • Empirical results on real data in Table 2 are not really better than existing work (for example, C-PCP).
  • Theoretical results are not very strong given existing proof technology (essentially, if there is a lot of data, it works). This is nice to have, but not particularly surprising given the method.
Questions for authors
  • The algorithm has many hyperparameters. How would you adapt them in practice for a new problem?
  • The section 3.4 on the residual refinement is a bit dense and results are almost the same in Table 2. Is it worth the effort?
  • For synthetic data (Table 1), the proposed method works significantly better than others, but not for real data (Table 2). Any explanation?
3 comments
AuthorsOpenReview ↗

Response to Reviewer tJZ8 (1/2)

We thank the reviewer for the encouraging assessment and constructive suggestions. We have expanded the main experiments to twenty shared splits, clarified the theoretical contributions, and improved the presentation of the empirical results and residual refinement.


W1. Improvements on real-world benchmarks

We appreciate the opportunity to clarify the real-data comparisons. LECP's main advantage is prediction-region efficiency, while its conditional-reliability diagnostics remain competitive rather than uniformly best.

We have expanded the main benchmark from ten to twenty shared splits. At α=0.10\alpha=0.10, the updated Table 2 gives the following root-volume comparisons between Res-LECP and C-PCP:

DatasetRes-LECPC-PCPReduction
Bio (2D)0.861.0719.6%
House (2D)0.861.6547.9%
Taxi (2D)2.493.4026.8%
Households (4D)2.522.9614.9%
RF1 (8D)0.320.4833.3%
RF2 (8D)0.310.4632.6%

The reductions are calculated from the displayed means. Thus, the efficiency improvement is present across all six real datasets, although its magnitude varies.

We have also added Figure 2 to make the efficiency–reliability trade-off more transparent. Averaged equally across the six datasets, Res-LECP achieves mean root volume 1.23, compared with 1.64 for CP2-PCP and 1.67 for C-PCP. CP2-PCP and C-PCP obtain better average ERT and WSC, respectively, but with appreciably larger prediction regions. We have clarified this trade-off in Section 5 rather than suggesting that LECP is uniformly best on every metric.

The updated paired analysis over twenty shared splits further supports the efficiency gains (Appendix E.3, Table 10). At α=0.10\alpha=0.10, Res-LECP achieves significantly smaller root volume than all six baselines on all six real datasets, giving 36/36 significant comparisons after Holm correction.


W2. Strength of the theoretical contributions

Finite-sample marginal coverage follows from standard split conformal calibration, as stated in Theorem 1. The additional theoretical analysis concerns the geometry and conditional coverage of the learned regions.

Under the stated estimation-consistency and regularity conditions, Theorem 2 establishes asymptotic conditional coverage on high-probability covariate sets and convergence in symmetric difference to the oracle conditional minimum-volume HDR. Appendix B.6 further connects the required estimation conditions to contrastive excess risk and conditional pinball regret.

We have added empirical analyses that examine these conclusions more directly. Appendix E.1 (Tables 8–9; Figure 9) evaluates symmetric difference with the oracle HDRs on four synthetic benchmarks. Both LECP variants outperform all six baselines, with 48/48 Holm-adjusted paired comparisons yielding p<0.001p<0.001. Appendix E.2 (Figure 10) further examines estimation and prediction-set errors at four sample sizes, showing decreasing mean errors between n=2500n=2500 and n=20000n=20000.

These results provide empirical support for the oracle-recovery mechanism established by the theory. Establishing finite-sample convergence rates for the practical neural training procedure remains an open problem.


Q1. Hyperparameter selection in practice

For a new dataset, we recommend starting with the two-scale local Gaussian proposal (relative standard deviations 0.5 and 1), M=1000M=1000, β=0.1\beta=0.1, and λ=0.01\lambda=0.01. When tuning is needed, λ\lambda can be selected from 0.01,0.1,1\\{0.01,0.1,1\\} using validation-calibrated volume.

Appendix D.2 reports sensitivity analyses and a new five-dataset validation study supporting this selection procedure. Moderate hyperparameter variations have limited impact in the tested settings; the main observed failure mode is an excessively narrow proposal, which can increase root volume by 98–257%. The final conformal calibration set is kept separate from all hyperparameter decisions.

AuthorsOpenReview ↗

Response to Reviewer tJZ8 (2/2)

Q2. Clarity and practical value of Res-LECP

Thank you for pointing out that Section 3.4 was difficult to follow. We have rewritten it around a simpler boosting-like interpretation: each residual stage learns an additive correction to the remaining log-density discrepancy of the current score. Boundary-focused weighting emphasizes the provisional prediction contour while preserving the ideal density ordering.

We agree that refinement offers limited gains on the main benchmarks. In the updated twenty-split experiments, validation selects the unrefined base score in 18/20 House splits and 15/20 Spiral splits. This suggests that additional residual stages are often unnecessary when the base LECP score is already accurate.

To investigate when refinement becomes more valuable, we constructed a controlled experiment from the two-dimensional Spiral benchmark. We generate Y=XγZY=X^\gamma Z, where ZZ follows the underlying conditional Spiral distribution. Increasing γ\gamma makes the conditional response scale vary more sharply across covariate values while retaining the underlying generating mechanism.

The newly added Table 7 (Appendix D.5) reports the following paired improvements in IoU with the oracle 90% HDR over ten seeds:

γ\gammaIoU gain (Res-LECP − LECP)95% CIHolm-adjusted pp
0+0.004[-0.004, 0.012]0.299
1+0.067[0.019, 0.115]0.024
2+0.145[0.068, 0.221]0.00623
3+0.277[0.228, 0.325]2.10×10−62.10\times10^{-6}
4+0.240[0.174, 0.305]6.94×10−56.94\times10^{-5}

The gains are statistically significant for γ=1,…,4\gamma=1,\ldots,4. At γ=4\gamma=4, the base LECP region has approximately 3.2 times the oracle HDR volume, compared with less than 1.7 times for Res-LECP, with comparable conditional reliability. Figures 7–8 further illustrate how refinement recovers boundaries closer to the oracle under strong scale heterogeneity.

These findings clarify when the added computation is worthwhile: Res-LECP can substantially improve geometric recovery when the base score retains systematic contour errors, while remaining optional on benchmarks where the base fit is already accurate.


Q3. Why are improvements larger on synthetic data?

The synthetic benchmarks deliberately test disconnected components, annular gaps, and thin curved supports, where learning flexible conditional HDR geometry can offer particularly large advantages. Real-world data may involve more intricate covariate–response relationships and noise, making the underlying conditional density geometry harder to estimate. Since the oracle HDRs of real datasets are unknown, however, we cannot determine precisely how much these factors contribute to the smaller relative gains.

Nevertheless, the improvements on real data remain meaningful and consistent. At α=0.10\alpha=0.10, LECP or Res-LECP attains the smallest mean root volume on all six real datasets, with statistically supported improvements as discussed above. Across α∈0.05,0.10,0.20\alpha\in\\{0.05,0.10,0.20\\}, an LECP variant achieves lower mean root volume than SLS in 17 of 18 real dataset–coverage-level comparisons (Appendix D.1).

We also believe it is important to interpret region size alongside conditional reliability. For example, at α=0.20\alpha=0.20 on RF1 and RF2, PCP produces slightly smaller root volumes (0.28 on both datasets), but its worst-slab coverage drops to 0.23 and 0.24, respectively, compared with 0.63 for Res-LECP on both. This illustrates why smaller prediction regions are not necessarily preferable when accompanied by substantial conditional undercoverage.

We appreciate the reviewer's thoughtful suggestions, which have helped us strengthen both the empirical evaluation and the presentation of our work.

Dear Reviewer tJZ8,

As the discussion period is drawing to a close, we wanted to briefly follow up on our responses and the updated manuscript. We have expanded the empirical evaluation and added further analyses to clarify the theoretical contributions and the practical value of residual refinement.

We would be grateful for any further thoughts you might have on these revisions, especially if there are any remaining questions or concerns.

Thank you again for your time and constructive feedback.

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

The paper proposes Lebesgue-contrastive prediction (LECP),which learns a score by logistic discrimination of observed responses against a Lebesgue reference measure. The learned score is closely related to the conditional density and is used as a nonconformity score to construct the prediction set via split conformal prediction. The coverage and consistency guarantees are established under some conditions. Numerical experiments show that their method tends to produce prediction sets with smaller volume.

Critical strengths
  1. The minimum-volume prediction set is characterized by the superlevel set of the conditional density. Building on this, the paper proposes a Lebesgue-constrastive learning method which is interesting.
  2. The paper studies finite-sample marginal coverage, asymptotic conditional coverage and convergence of the estimated set to the oracle minimum-volume set.
  3. Extensive experiments are provided to demonstrate the performance of the proposed method.
Critical weaknesses
  1. The idea of using contrastive learning to estimate the conditional density is not new, as also mentioned by the authors (e.g., CINDES: Classification induced neural density estimator and simulator). The key idea is to choose a reference measure whose density does not depend on yy, so that the density ratio is related to the target density. This reduces the novelty of the proposed contrastive formulation.
  2. In theoretical analysis, estimation consistency is assumed directly rather than established for the proposed estimator.
  3. The motivation for Res-LECP is not sufficiently clear, as the numerical results show only limited improvement over the base LECP method.
Questions for authors
  1. What are the main advantages of the proposed estimator over existing density- or density-ratio-based methods?
  2. The authors should discuss under what conditions the consistency assumptions (Assumption 3) are expected to hold. In the synthetic experiments, the errors may be provided to provide empirical support.
  3. It would be helpful to clarify when residual refinement is expected to provide substantial benefits.
  4. The choice of proposal density may be nontrivial in practice. How should it be selected in real applications? Moreover, different datasets appear to use different default hyperparameter settings, which raises concerns about the consistency and robustness of the tuning procedure across datasets.
  5. How does the proposed method scale with the dimensions of the covariates and responses?
  6. Ten random seeds may be insufficient to assess the stability of the results. More repetitions would help provide a more reliable comparison. For the synthetic experiments, since the oracle prediction sets are available, it would be informative to directly quantify the discrepancy between the estimated and oracle sets, for example using symmetric difference and provide corresponding visualizations.
4 comments
AuthorsOpenReview ↗

Response to Reviewer Jtpx (1/3)

We thank the reviewer for the thoughtful assessment. We have clarified the method, added oracle-recovery and estimation-error analyses, and expanded the main benchmark to twenty shared splits in response to Q6.


W1 / Q1. Novelty and advantages over existing density estimators

Classification-based conditional density estimation is established, including in NCE and CINDES. A suitably weighted uniform reference on a bounded domain can recover the same density ordering as LECP. Our contribution lies in how this ordering is estimated and used for conditional minimum-volume prediction.

CINDES constructs a neural conditional density estimator by distinguishing observed responses from samples drawn from a specified global reference distribution. For a general reference density q(y)q(y), the resulting classification ratio depends on q(y)/p(y∣x)q(y)/p(y\mid x), so recovering the target density ordering requires accounting for the reference density.

LECP instead fixes its population reference to the Lebesgue intensity λμ\lambda\mu, while allowing negative samples to be drawn from observation-dependent local proposals qiq_i. Importance correction ensures that

\left[\frac{\ell_\beta^-(s(x,Y^-))}{q_i(Y^-)}\right] =\int_{\mathcal Y}\ell_\beta^-(s(x,y)) dy.$$ Thus, local sampling changes the Monte Carlo approximation but not the population objective, whose minimizer remains $$s^\star(x,y)=\beta\log\frac{\lambda}{p(y\mid x)}.$$ This separation allows reference samples to concentrate near observed responses while preserving the density ordering needed for conditional HDR recovery. Appendix D.3 examines the practical importance of this design. Removing both locality and importance correction increases root volume by **39.9–94.6%** across five datasets. A global-reference classifier with post-hoc density correction retains **33.4–76.8%** excess root volume on House and RF1. Removing locality alone increases root volume by **70.9%** on RF1 (8D), although its effect is much smaller on the lower-dimensional datasets tested. These are our controlled mechanism ablations. LECP also uses conditional-quantile tracking to locate the contour and independent split conformal calibration for marginal validity. Compared with flow-based approaches, it requires neither an invertible density model nor explicit density normalization for score learning or prediction-set membership evaluation. Compared with SLS, it learns conditional density geometry through a contrastive objective instead of directly differentiating through a volume surrogate with a specified frontier parameterization. The updated runtime experiment (Appendix D.4) shows that full-test prediction-set membership evaluation is **2.0–8.8 times faster** than L-CP, the fastest evaluated baseline. We have revised Section 2 to explain these methodological connections and differences more clearly. --- ### **W2 / Q2. Assumption 3 and estimation consistency** The consistency conditions in Assumption 3 concern the final fitted score and conditional level. We do not establish them unconditionally for the practical neural training algorithm, but Appendix B.6 provides sufficient conditions in terms of the population learning objectives. Specifically, Proposition 2 shows that contrastive excess risk controls the data-law and Lebesgue-domain discriminator errors: $$\varepsilon_{D,n} \leq \frac{\mathcal R(\widehat s_n)-\mathcal R^\star}{2\beta}, \qquad \varepsilon_{\mu,n} \leq \frac{\mathcal R(\widehat s_n)-\mathcal R^\star}{2\beta\lambda}.$$ Proposition 3 further establishes level consistency from vanishing conditional pinball regret, given score consistency and the stated regularity conditions. Together, these results connect the estimation requirements in Theorem 2 to the population contrastive and pinball objectives. Establishing finite-sample convergence rates for the implemented neural optimizer remains an open question. We added **Appendix E.2 (Figure 10)**, evaluating all three estimation errors in Assumption 3 at $n=2{,}500$, $5{,}000$, $10{,}000$, and $20{,}000$, using ten nested splits on each of the four synthetic benchmarks. All three mean errors decrease between the smallest and largest sample sizes on every dataset, as does the oracle symmetric difference: | Dataset | $n=2{,}500$ | $n=20{,}000$ | |---|---:|---:| | Comb | 0.270 | 0.124 | | Rings | 0.649 | 0.106 | | Spiral | 2.440 | 0.844 | | Lemniscate | 0.0289 | 0.0157 | The Lebesgue-domain error is evaluated on specified finite integration windows. These findings provide empirical support for the recovery mechanism in Theorem 2.
AuthorsOpenReview ↗

Response to Reviewer Jtpx (2/3)

W3 / Q3. Motivation and benefits of Res-LECP

We have rewritten Section 3.4 to explain Res-LECP as a boosting-like correction of remaining log-density error.

At each stage, the current cumulative score defines an unnormalized density surrogate p~b−1(y∣x)=exp⁡−s[b−1](x,y)/β\widetilde p_{b-1}(y\mid x)=\exp\\{-s^{[b-1]}(x,y)/\beta\\}. A residual network is trained by weighted logistic contrast against a reference constructed from this surrogate. At the population optimum,

=-\beta\log p(y\mid x)+a_b(x).$$ The boundary-focused weight appears in both contrast measures and cancels from their density ratio, preserving the ideal density ordering. The $x$-dependent offset is absorbed by subsequent conditional-level estimation. The updated main experiments confirm that refinement is often unnecessary when the base fit is already accurate: validation retains the unrefined model in **18/20 House splits and 15/20 Spiral splits**. To identify when refinement becomes more useful, Appendix D.5 studies a controlled transformation of the two-dimensional Spiral benchmark, $Y=X^\gamma Z$, where increasing $\gamma$ amplifies covariate-dependent response-scale heterogeneity while retaining the underlying Spiral structure. The new **Table 7** reports paired oracle HDR-IoU improvements of **+0.067, +0.145, +0.277, and +0.240** for $\gamma=1,2,3,4$, respectively, all significant after Holm correction. The largest gain, at $\gamma=3$, has a paired 95% CI of **[0.228, 0.325]**. At $\gamma=4$, LECP's relative prediction volume reaches approximately 3.2 times the oracle, compared with less than 1.7 times for Res-LECP, with comparable conditional reliability. These results identify a setting in which residual refinement has a substantial benefit, while supporting its role as an optional refinement when the base score is already sufficiently accurate. --- ### **Q4. Proposal choice and hyperparameter tuning** Our default proposal is an equally weighted mixture of two diagonal Gaussians centered at each observed response, with relative standard deviations 0.5 and 1. The narrower component concentrates samples near observations, while the wider component provides additional tail exploration. We use $M=1000$ reference samples, $\beta=0.1$, and a default reference intensity of $\lambda=0.01$. For a new dataset, we recommend starting with these settings and selecting $\lambda$ from $\\{0.01,0.1,1\\}$ using validation-calibrated volume when tuning is needed. Proposal scale also deserves attention: Appendix D.2 shows that reducing it to 0.1 times the default increases root volume by **98–257%**. In contrast, moderate variations of $\lambda$, $\beta$, and $M$ have limited impact in the tested settings, and increasing $M$ from 1000 to 2000 offers little additional benefit relative to its computational cost. Regarding the different reference intensities, our main experiments use $\lambda=1$ on RF1/RF2 following validation selection on a split, and the default $\lambda=0.01$ elsewhere. To examine whether a common tuning rule can account for these choices, we additionally evaluated validation-based selection from $\\{0.01,0.1,1\\}$ on five datasets with five splits each, keeping other hyperparameters fixed. The procedure selected 0.01 for House and 1 for RF1 in all five splits (Appendix D.2). All hyperparameter choices are made without using the final conformal calibration set. --- ### **Q5. Scaling with covariate and response dimensions** Our benchmarks cover up to **64 covariates and 8 response dimensions**. LECP uses ordinary scalar-output networks and requires no invertible density parameterization. Prediction-set membership can be evaluated using the score and conditional-level networks, without sampling from a fitted generative model at inference. We updated the runtime comparison to twenty shared splits (Appendix D.4, Figure 6). Full-test membership evaluation is **2.0–8.8 times faster** than L-CP, the fastest baseline. The trade-off is training cost: LECP takes **3.3–20.2 times longer** than fitting the shared MAF, primarily because of repeated Monte Carlo reference sampling and validation-volume evaluation. As response dimension increases, adequate proposal coverage and the variance of importance weights become more challenging. The locality ablation on RF1 (8D), where removing locality increases root volume by 70.9%, illustrates the importance of proposal design. More efficient adaptive sampling strategies for higher-dimensional responses remain an important direction for future work.
AuthorsOpenReview ↗

Response to Reviewer Jtpx (3/3)

Q6. Statistical stability and oracle comparisons

Thank you for suggesting additional repetitions and direct comparisons with the oracle prediction sets. We have now expanded the entire main benchmark from 10 to 20 shared random splits, covering all ten datasets at α=0.05,0.10,0.20\alpha=0.05,0.10,0.20. Tables 1, 2, 4, and 5 have been updated accordingly.

We also recomputed the paired statistical analyses over twenty splits (Appendix E.3, Table 10). At α=0.10\alpha=0.10, Res-LECP achieves significantly smaller root volume than all six baselines on all six real datasets (36/36 comparisons), while LECP does so in 30/36 comparisons. We report pointwise 95% confidence intervals and two-sided paired tests, with Holm correction across the six baselines within each dataset, miscoverage level, and LECP variant.

For direct oracle comparison, we added Appendix E.1 (Tables 8–9; Figure 9). Table 8 evaluates the symmetric difference (SymDiff) between calibrated regions and oracle conditional HDRs on all four synthetic datasets. The complete results at α=0.10\alpha=0.10 are given below (mean (standard deviation) over ten shared splits; smaller is better):

MethodCombRingsSpiralLemniscate
LECP0.129 (0.023)0.139 (0.027)1.328 (0.319)0.0184 (0.0022)
Res-LECP0.127 (0.017)0.137 (0.023)1.317 (0.327)0.0168 (0.0031)
SLS0.852 (0.026)0.833 (0.019)2.221 (0.358)0.1269 (0.0701)
L-CP2.961 (0.050)1.032 (0.042)7.193 (0.831)0.1687 (0.0169)
C-HDR3.036 (0.048)0.951 (0.024)5.899 (0.696)0.1293 (0.0190)
C-PCP4.231 (0.113)1.615 (0.065)9.876 (0.417)0.2280 (0.0196)
CP2-PCP4.210 (0.114)1.588 (0.060)9.809 (0.430)0.2256 (0.0201)
PCP3.975 (0.085)1.578 (0.052)9.902 (0.371)0.2244 (0.0199)

Both LECP variants outperform all six baselines on all four benchmarks, with all 48 paired comparisons yielding Holm-adjusted p<0.001p<0.001. Figure 9 provides corresponding visualizations for all eight methods, illustrating the recovery of disconnected components and curved HDR boundaries.

The oracle comparison uses ten shared splits, whereas the expanded main benchmark and root-volume significance analysis use twenty. Together, these additions provide stronger replication evidence and directly quantify the geometric recovery requested by the reviewer.

We appreciate the reviewer's suggestions, particularly the request for more repetitions and oracle-set comparisons. They have helped us strengthen the empirical evaluation and clarify the conditions under which LECP is expected to recover the desired prediction regions.

Dear Reviewer Jtpx,

As the discussion period is approaching its end, we wanted to briefly follow up on our responses and the revised manuscript. We have added further theoretical clarifications and experimental analyses to address your questions, including expanded evaluations over 20 splits and direct comparisons with oracle prediction regions.

If you have had a chance to review these updates, we would greatly appreciate any further feedback, particularly if any concerns remain.

Thank you again for your time and thoughtful review.

Reviewer_oX84OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes LECP, a method for learning nonconformity scores for multivariate conformal regression. It trains a scoring model by contrasting observed responses with sampled reference outputs, then combines the learned score with an input-dependent threshold and split conformal calibration to construct flexible prediction regions.

Critical strengths
  1. The learned scoring function enables flexible prediction regions and improves volume efficiency while maintaining competitive coverage diagnostics.
  2. The synthetic and real-data experiments generally support the claimed efficiency gains and demonstrate the method’s ability to capture complex, disconnected prediction regions.
Critical weaknesses
  1. The novelty of the overall score-learning framework needs clearer positioning. Prior work, such as Uncertainty Quantification over Graph with Conformalized Graph Neural Networks (Huang et al., 2023), already learns a correction to nonconformity scores before conformal calibration to improve efficiency. LECP follows a related learn-then-calibrate strategy for multivariate regression. Although its Lebesgue-contrastive objective differs, the authors should clarify the methodological contribution beyond adapting this established strategy to multivariate outputs.
  2. Res-LECP provides limited improvements over the base method on the main benchmarks. Many reported results are identical or nearly identical, leaving the practical value of the additional residual stages unclear relative to their training complexity.
Questions for authors

See weakness.

9 comments
AuthorsOpenReview ↗

Response to Reviewer oX84 (1/2)

We thank the reviewer for recognizing the flexibility of LECP's prediction regions and its empirical efficiency gains. We have revised the related-work discussion, clarified the motivation for residual refinement, and completed additional experiments that help explain when the refinement is beneficial.


W1. Novelty relative to CF-GNN and the learn-then-calibrate framework

We appreciate the comparison with CF-GNN. Learning or correcting predictions before conformal calibration is indeed an established approach. The distinction we aim to make clearer is what the learning objective recovers and how the resulting score is estimated.

CF-GNN uses graph-neighborhood information to correct pretrained GNN predictions, with the correction trained by minimizing a differentiable surrogate of conformal inefficiency. LECP instead learns a scalar field sθ(x,y)s_\theta(x,y) directly over the multivariate response space through a Lebesgue-contrastive objective. Its population minimizer satisfies

s⋆(x,y)=βlog⁡λp(y∣x).s^\star(x,y)=\beta\log\frac{\lambda}{p(y\mid x)}.

Thus, the score's sublevel sets coincide with conditional density superlevel sets, recovering the geometry of the conditional minimum-volume HDRs. CF-GNN directly optimizes a surrogate of prediction-set efficiency, whereas LECP learns the conditional density ordering that determines minimum-volume regions. The contrastive score objective itself is independent of the coverage level, although conditional-level fitting and checkpoint selection remain coverage-dependent.

Another important aspect is the separation of the population reference measure from the sampling proposal. LECP fixes its population objective against the Lebesgue intensity λμ\lambda\mu but uses observation-centered local proposals, corrected by importance weights, to estimate the reference integral. This concentrates negative samples near observed responses without changing the population density-ordering target. The classification-to-density principle itself is well established; our contribution is to use this estimation design specifically for learning conditional HDR geometry.

The mechanism ablations in Appendix D.3 provide evidence for its practical value. Removing importance correction increases root volume by 11.3–38.8% on the tested real datasets. Removing both locality and importance correction increases it by 39.9–94.6% across five benchmarks, while removing locality alone leads to a 70.9% increase on the eight-dimensional RF1 dataset.

LECP then combines the learned density-ordering score with conditional-quantile tracking, validation-based score–level selection, and independent conformal calibration. Unlike direct volume optimization such as SLS, the score-learning loss does not require differentiating through prediction-set volume or specifying a flow-based boundary parameterization.

We have clarified these distinctions from CF-GNN, CINDES, and SLS in Section 2. We hope this makes clearer the methodological contribution beyond the general learn-then-calibrate strategy.

AuthorsOpenReview ↗

Response to Reviewer oX84 (2/2)

W2. Practical value of Res-LECP

We agree that Res-LECP offers only modest improvements on most main benchmarks. The expanded 20-split experiments help explain why: validation selects the unrefined base score in 18/20 House splits and 15/20 Spiral splits (Appendix D.5). When the base score already captures the relevant conditional geometry, additional residual stages offer limited benefit.

To investigate when refinement becomes useful, we constructed a controlled stress test based on the two-dimensional Spiral benchmark. We draw X∼Unif[0.25,1]X\sim\mathrm{Unif}[0.25,1], generate a Spiral response ZZ conditional on a rescaled version of XX, and set

Y=XγZ,γ∈0,1,2,3,4.Y=X^\gamma Z,\qquad \gamma\in\\{0,1,2,3,4\\}.

Increasing γ\gamma creates progressively stronger variation in response scale across covariate values while retaining the underlying conditional Spiral structure. The same underlying (X,Z)(X,Z) draws are used across the five settings, allowing us to isolate the effect of increasing scale heterogeneity.

The new Table 7 (Appendix D.5) reports the paired improvement in IoU with the oracle 90% HDR:

γ\gammaIoU gain (Res-LECP − LECP)95% CIHolm-adjusted pp
0+0.004[-0.004, 0.012]0.299
1+0.067[0.019, 0.115]0.024
2+0.145[0.068, 0.221]0.00623
3+0.277[0.228, 0.325]2.10×10−62.10\times10^{-6}
4+0.240[0.174, 0.305]6.94×10−56.94\times10^{-5}

These results use ten paired seeds, with pointwise 95% confidence intervals and Holm correction across the five γ\gamma values. At γ=0\gamma=0, the two methods are nearly indistinguishable. For γ=1,…,4\gamma=1,\ldots,4, the improvements are statistically significant, with the largest gain occurring at γ=3\gamma=3.

The improvement is also reflected in prediction-region size. At γ=4\gamma=4, LECP's region has approximately 3.2 times the oracle HDR volume, whereas Res-LECP remains below 1.7 times, with comparable conditional reliability. Figures 7–8 show how the base regions increasingly extend beyond the oracle contour, while residual refinement yields a closer geometric fit.

We have rewritten Section 3.4 to clarify the mechanism. Each residual stage learns an additive correction to the current score's remaining log-density discrepancy. Boundary-focused weighting emphasizes the provisional prediction contour while preserving the ideal density ordering; an xx-dependent normalization offset is subsequently absorbed by conditional-level estimation.

Taken together, these results suggest that Res-LECP is particularly useful when substantial contour errors remain in the base score, rather than being necessary for every dataset. We therefore present it as an optional refinement whose additional training cost is weighed against validation-estimated improvements.

We appreciate the reviewer's careful consideration of both novelty and practical complexity. If there are further specific concerns about the method's soundness, we would welcome the opportunity to address them.

Thank you for the additional experiments. Could you confirm whether you have compared training with and without importance correction across different proposal scales? I found the separate sensitivity and ablation analyses in Appendices D.2–D.3, but not this joint comparison. If it is not included, please consider adding it with matched sampling and training budgets to clarify how the benefit of importance weighting depends on sampling locality.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions

We thank all reviewers for their thoughtful feedback. We have completed the additional experiments and uploaded an updated manuscript, with revisions marked in blue. The main changes are summarized below.

1. Unified volume evaluation (Section 5; Appendix C.3). All eight methods, including SLS, are now evaluated on the full test sets using the same importance-sampling volume estimator and shared draws from the fitted MAF. For the eight-dimensional RF datasets, we increased the number of draws per test covariate from 5,000 to 10,000. We also clarified the efficiency metric as root volume (Root-Vol), the dd-th root of the mean test-set prediction volume.

2. Expanded experiments and statistical comparisons (Section 5; Appendices D.1 and E.3). Following the request for more repetitions, we expanded the main benchmark from 10 to 20 shared random splits across all ten datasets and three miscoverage levels, updating Tables 1, 2, 4, and 5. Appendix E.3 (Table 10) now reports paired 95% confidence intervals and Holm-adjusted tests over 20 splits. At α=0.10\alpha=0.10, LECP or Res-LECP attains the smallest mean root volume on all ten datasets. On the six real datasets, Res-LECP and LECP significantly outperform the baselines in 36/36 and 30/36 comparisons, respectively.

3. Methodological clarity (Sections 2 and 3.4). We expanded the discussion of CF-GNN, CINDES, and SLS to clarify LECP's density-ordering objective and locally importance-sampled estimator. We also rewrote Section 3.4 to explain Res-LECP's stagewise log-density correction, boundary weighting, and validation-based residual-prefix selection.

4. Improved visualizations and computational comparisons (Section 5; Appendix D.4). We refined the synthetic prediction-region visualizations and expanded the Taxi comparison to all eight methods. The new Figure 2 illustrates the trade-off between root volume and conditional-reliability diagnostics across six real datasets. We also updated the runtime analysis to 20 splits, reporting model-fitting and full-test membership-evaluation times separately, with 95% confidence intervals.

5. Oracle recovery and estimation errors (Appendices E.1–E.2). We added direct symmetric-difference (SymDiff) comparisons with oracle conditional HDRs, paired significance tests, and visualizations of all eight methods on four synthetic benchmarks (Tables 8–9; Figure 9). We also examined discriminator errors, conditional-level errors, prediction-set symmetric difference, and CC-MAE across four sample sizes (Figure 10). These studies use ten shared or nested splits.

6. Hyperparameter selection and ablations (Appendices D.2–D.3). We evaluated a common validation-based rule for selecting the reference intensity λ\lambda on five datasets, using an independent split for final calibration. We also corrected the reported percentage changes and clarified the effects of proposal locality and importance correction.

7. When residual refinement helps (Appendix D.5). We added Table 7 with paired HDR-IoU gains, 95% confidence intervals, and Holm-adjusted tests across five levels of conditional-scale heterogeneity. Res-LECP significantly improves oracle HDR IoU for γ=1,…,4\gamma=1,\ldots,4, with the largest gain at γ=3\gamma=3: +0.277+0.277 (95% CI: [0.228,0.325][0.228,0.325]). This controlled study uses ten paired seeds; the updated main-benchmark residual-prefix frequencies are based on 20 splits.

8. Further analysis of importance correction (Appendix E.4). We added a joint study of importance correction and proposal scale across five datasets and five scales, using matched sampling and training budgets. The results reveal dataset-dependent interactions between sampling locality and the efficiency gains from importance correction. We further characterize the uncorrected objective as learning a density ratio against a smoothed conditional reference, and explain how the resulting density ordering affects HDR volume.

9. Proposal centering versus bandwidth (Appendix E.5; Figure 12). We added a comparison of observation-centered and fixed-center proposals across five scales and five datasets, using five paired splits and matched sampling and training budgets. Both variants use importance correction. The differences diminish at broader bandwidths on several datasets, while observation centering retains a significant root-volume advantage on RF1 at four of the five scales.

Detailed responses to each review are provided separately. We appreciate the reviewers' suggestions, which have helped us strengthen the presentation, evaluation consistency, and empirical evidence of the paper.

AuthorsOpenReview ↗

Closing Summary of Revisions and Discussion

Dear Area Chair and Reviewers,

As the discussion period draws to a close, we would like to thank the reviewers for their constructive comments and follow-up questions. We have provided point-by-point responses to all concerns and questions raised in the three reviews and substantially revised the manuscript, with changes marked in blue.

Our revisions include an expanded evaluation over 20 shared splits across ten datasets and three miscoverage levels, unified full-test volume estimation, paired statistical comparisons, and direct comparisons with oracle conditional HDRs. At α=0.10\alpha=0.10, LECP or Res-LECP achieves the smallest mean root volume on all ten benchmarks. On the six real datasets, Res-LECP achieves significantly smaller root volumes than all six baselines in all 36 paired comparisons after Holm correction. The new oracle comparisons further show significantly smaller symmetric differences for both LECP variants across all four synthetic benchmarks.

We have also strengthened the theoretical and methodological discussion, clarified the distinctions from existing contrastive density estimation and conformal methods, and connected the estimation-consistency assumptions to population excess risks. Additional analyses address hyperparameter selection, computational costs, and the practical role of residual refinement, including a controlled experiment identifying settings where Res-LECP provides substantial improvements.

In response to the follow-up discussion, we further added Appendices E.4 and E.5, examining importance correction, proposal bandwidth, and observation-dependent centering through controlled comparisons with matched sampling and training budgets. These analyses clarify the distinct roles of importance correction and proposal locality, including their dataset- and bandwidth-dependent effects on finite-sample efficiency.

The complete revisions are summarized in our earlier Summary of Revisions, with detailed responses and supporting results provided in the reviewer-specific threads. We hope the revised manuscript and additional evidence will be carefully considered in the final assessment.

Thank you again for your time and consideration.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.