acceptodds
Under review as a conference paper at ICLR 2027

CoCoNet: Convex Conformal Network for Multi-dimensional Uncertainty Regions

Abstract

Conformal prediction (CP) provides distribution-free coverage guarantees under exchangeability, but the efficiency of its prediction regions depends critically on the choice of the conformity score. This is particularly important in multivariate regression, where dependencies across output dimensions induce complex joint uncertainty structures. This paper introduces CoCoNet (***Co***nvex ***Co***nformal ***Net***work) for learning convex conformity scores that induce flexible, data-adaptive prediction regions. CoCoNet learns the score from held-out data by optimizing prediction-region efficiency while enforcing convexity with respect to the residual. CoCoNet then applies split-conformal calibration to retain finite-sample coverage guarantees. The paper presents two variants of the method: CoCoNet, based on piecewise-linear scores, and CoCoNet, based on Input Convex Neural Networks. Experiments compare CoCoNet with other CP baselines over synthetic and real-world datasets with output dimensions ranging from 2 to 16. The results demonstrate the flexibility of CoCoNet in learning different response types, with CoCoNet ranking first on 25 tasks out of 30 in terms of efficiency.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conformal Prediction, Uncertainty Quantification, Multivariate Prediction Regions, Learned Conformity Scores, Input-Convex Neural Networks, Adaptive Prediction Sets, Distribution-Free Prediction

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

This paper studies multivariate regression. The broader issue is that different functions used in conformal prediction can lead to substantial differences in the sizes and shapes of prediction regions. Within this setting, the paper focuses on constructing convex conformal prediction regions for multidimensional regression. Other approaches restrict the available regions to a small number of predefined geometric families, which constrains the resulting regions and may produce unnecessarily large prediction regions. To address this issue, the paper proposes CoCoNet (Convex Conformal Network) and reformulates the problem in terms of score functions. It develops variants based on different scoring schemes and conducts extensive experiments across regression tasks to demonstrate the effectiveness of the approach.

Critical strengths

• The logic of the method derivation is strict and complete, and the definition of the problem is clear. • The experimental results are rich and convincing, and the performance shown in the experiments is very good. • The achievements in this paper are rich, and the amount of work is large.

Critical weaknesses

• The structure of the article is severely unbalanced. The main body cannot support the whole logic of the research, especially because the “Related Work” section is included in the supplementary materials, which makes the importance of the contributions hard to evaluate and understand. Some concepts lack explanation in the main body; their explanations are all provided in the appendices, requiring readers to search for them themselves.

• The background and motivations are vague in the abstract and introduction, making it hard for me to evaluate the value of the work from the starting point.

• The abstract and introduction seem to consist of key points summarized from the appendices and “combined” into content in a sequential order rather than through logical connections. This makes me confused and harder to understand the content, especially because some main mechanism of the improvement is not illustrated in the main body.

• The logic of the reason why the method can handle the problem is weak: In Section 3, Eqs. (9)–(12) optimize the prediction-region size at the fixed training threshold of 1, whereas the final region is defined using a separately calibrated threshold. Reducing the former does not automatically reduce the latter at the target coverage level. The coverage and miss penalties may help bridge this gap, but the paper does not isolate their contribution: the ablation in Appendix E.1 focuses on input dependence and the contextual encoder. The reported post-calibration results provide empirical evidence of overall performance, but do not by themselves establish how the individual loss terms preserve the training-time gains after calibration. Reporting the calibration-induced change in region size and conducting a loss-term ablation would clarify this connection. This concern relates to the efficiency objective, not the validity of independent split-conformal calibration.

• The correlation between the definition and the capability of the network is weak : Proposition 3 establishes the existence of an admissible convex score representing a compact, full-dimensional convex region containing the point prediction in its interior. This is a result for the unrestricted admissible function class, not an approximation or learning guarantee for the particular networks used in the experiments. The Spline implementation still produces rotated hyperrectangles at a fixed threshold (Eq. (17)), and the ICNN implementation has finite width and depth and includes a positive quadratic term (Eq. (21); Table 5). Therefore, the existence result alone does not establish that these implementations can represent or learn arbitrary convex geometries. The paper should distinguish representation by the abstract score class, approximation by the chosen architecture, and recovery through training. The synthetic examples support specific geometric cases, but do not establish the unrestricted claim. An architecture-specific justification, or a narrower statement of the claim, is needed.

• The table in the experimental section is extremely unfriendly to human reviewers. The characters are too small, and the results are redundant, making it hard to find the key performance results.

Questions for authors

Question 1 : Could the authors provide clearer application contexts, background, and motivation, as well as clearer and more direct logical connections between the background, motivation, problem, and proposed method?

Question 2 : Could you provide a concise explanation of the methodological contribution relative to the closest prior approaches, and explain how the proposed volume-minimization objective is computed and differentiated? These points are central to evaluating the paper, but the main text does not sufficiently connect them, requiring the reader to reconstruct the explanation from material scattered across the manuscript and appendices.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Usyg (1/3)

We appreciate your feedback, and we are glad that you found the derivation strict and complete and the experimental results rich and convincing. We provide point-by-point responses below (in three parts); all changes in the revised PDF are marked in blue.

Structure, background and motivation (Question 1)

Thank you for these suggestions. We revised the front of the paper along these lines:

  • Abstract. It now names the mechanism of the improvement: an input-dependent convex score learned end to end by directly minimizing the size of its sublevel sets, followed by split-conformal calibration.
  • Motivation. The convexity paragraph of the introduction is rewritten around concrete application contexts, robust optimization, planning and control, where convex uncertainty sets are the standard requirement. It gives three reasons: convexity acts as geometric regularization (no fragmented or disconnected regions in data-sparse areas); membership testing, projection and the "must hold for every point in the set" constraints are tractable over convex sets via duality; and for constraints convex in the outcome, the worst case over any set equals the worst case over its convex hull, so a smaller non-convex region gives no better decision.
  • Positioning. A new paragraph in the introduction covers the two closest works, MVCS and Yeh et al., and states what CoCoNet does differently.
  • Mechanism in the main text. The caption of Figure 1 now walks through the framework step by step and connects each component to where it is developed: the encoder to Section 3 and its training details, the convex score to the admissibility and convexity results, the volume objective to the derivation of its computation and gradients, and calibration to the coverage guarantee. Section 3 also now summarizes how the volume objective is computed, with the full derivation of the volume estimation and its differentiation in Appendices C.2 (spline variants) and C.3 (CoCoNet-ICNN); Part 3 gives a short overview.

The full related work remains in Appendix A because of the page limit, but its main contrast is now in the introduction. We would be grateful to know of any passage that remains unclear, so that we can expand it further in the main text.

Regarding the results table

"The table in the experimental section is extremely unfriendly to human reviewers. The characters are too small, and the results are redundant."

Thank you for pointing this out. Table 2 now has one number per cell, the size over ten seeds with its standard deviation; the best entry is in blue, the second underlined, and a mean-rank row is at the bottom. The sizes are dimension-normalized, (mean volume)1/p(\text{mean volume})^{1/p}, which keeps tasks of different dimension on one scale. The empirical coverage, about 90% for every method, moved to Appendix Table 9, and the comparison with all 15 methods, including the non-convex baselines, is in Appendix Table 8, with a rank figure (Figure 4).

AuthorsOpenReview ↗

Response to Reviewer Usyg (2/3)

Regarding the training threshold and the coverage and miss penalties

"Reducing the former does not automatically reduce the latter at the target coverage level. The coverage and miss penalties may help bridge this gap, but the paper does not isolate their contribution."

Thank you for this careful observation. As you note, minimizing the size at τ = 1 does not by itself minimize the calibrated region. The revised Section 3 now states that the training threshold is fixed at τ = 1 to simplify the optimization: the coverage terms push the unit sublevel set towards coverage 1 − α, which keeps the size objective from shrinking the set far below the target coverage, and calibration then adjusts its level. The new Appendix E.4 (Table 13) reports the calibration-induced change in the threshold: over all tasks and seeds, the median q̂ is 1.68 for CoCoNet-ICNN, 1.80 for CoCoNet-Spline-ℓ∞ and 1.98 for CoCoNet-Spline-ℓ1. All reported sizes are those of the calibrated regions.

Following your suggestion, the new loss-term ablation (Appendix E.7, Table 20) isolates the two penalties: it compares the full objective with training after removing the aggregate coverage term L_cov or the instance-level miss term L_miss (the size term is kept in every variant, the remaining weight is selected on validation, five seeds). Calibrated size at 90% coverage (lower is better):

DatasetpFull objectiveWithout L_covWithout L_miss
ENB27.557.718.55
Bias-Corr. Temp.23.883.914.02
Online Retail II22.022.331.97
SCPF35.165.896.50
QM841.101.101.19
Communities Crime65.946.046.42
QM9-8D84.944.945.23
Knee Contact Force126.656.737.24
QM7b141.781.811.85
OES97162.873.657.62
Mean rank ↓1.202.002.80

The full objective gives the smallest calibrated region on eight of the ten datasets. Removing L_miss has the larger effect: the region grows on nine of the ten datasets, by 4–13% on most of them, 26% on SCPF and a factor of 2.7 on OES97. Removing L_cov costs less, but still up to 27% (OES97). The penalties therefore do not create the efficiency by themselves; they keep the trained set close to the target coverage, so that the training-time gains survive calibration.

Regarding the expressiveness claim

"The existence result alone does not establish that these implementations can represent or learn arbitrary convex geometries. An architecture-specific justification, or a narrower statement of the claim, is needed."

Thank you for this careful distinction; we have narrowed the claim accordingly. The revised Section 3 separates the abstract score class from the architectures: Proposition 3 shows that, within the unrestricted admissible score class, every compact, full-dimensional convex region containing the point prediction is a score sublevel set, but a given spline or ICNN represents only part of this class. The variants offer increasing flexibility:

  • CoCoNet-Spline-ℓ∞: rotated, asymmetric boxes.
  • CoCoNet-Spline-ℓ1 (new): rotated, asymmetric polytopes.
  • CoCoNet-ICNN: smooth convex regions of flexible shape; its positive quadratic term and smooth activations make the regions smooth, so it rounds the corners of polygonal targets.

None of them represents every convex region, and the paper does not claim an approximation or recovery guarantee for the trained networks. The evidence at that level is empirical, through the synthetic targets with known optimal regions (Table 1, IoU and V/V*).

AuthorsOpenReview ↗

Response to Reviewer Usyg (3/3)

Regarding the contribution and the volume objective (Question 2)

"Could you provide a concise explanation of the methodological contribution relative to the closest prior approaches, and explain how the proposed volume-minimization objective is computed and differentiated?"

Contribution. Prior learned conformal regions for multivariate regression come in two kinds: a fixed convex family whose parameters are learned (boxes, ellipsoids, the norm balls of MVCS [1]), or sets that are not convex in general (OT-CP, CONTRA, JAPAN, PBS). CoCoNet learns the convex shape itself, as an input-dependent score whose sublevel sets are convex by construction, trained by directly minimizing their size and then calibrated by split conformal prediction. It is a framework with different degrees of flexibility and structure: the spline variants produce efficient structured sets (rotated asymmetric boxes and polytopes), and CoCoNet-ICNN produces smooth, flexible convex sets. Relative to the closest works:

  • MVCS [1] also minimizes region volume, but over norm balls, so its input-adaptive regions are symmetric about the point prediction; its asymmetric multi-norm variant is used as one global shape. CoCoNet learns input-adaptive, asymmetric convex regions.
  • Yeh et al. [2] also learn input-dependent convex sets with an input-convex network, but they train for the performance of one downstream decision problem, which must be known in advance, and their ReLU network gives polytopes. CoCoNet minimizes the region size directly, so the same regions serve any downstream problem, and it covers both polytopes and smoothly curved regions. It also works on top of any pretrained point predictor.

Volume computation (summarized in Section 3; full derivation in Appendices C.2–C.3). For CoCoNet-Spline-ℓ∞, the region is a rotated box whose volume is the product of its side lengths; the CoCoNet-Spline-ℓ1 polytope also has a closed-form volume. The CoCoNet-ICNN region is convex and contains the point prediction, so its volume follows from its radial boundary: V = vₚ · 𝔼ᵤ[ρ(u)ᵖ], where vₚ is the volume of the unit ball and ρ(u) is the radius at which the score reaches the threshold along direction u. Because the score is strongly convex and vanishes at zero residual, this radius exists and is unique, and it is found by a one-dimensional root search. The directions are equally spaced for p = 2 and scrambled Sobol points for p > 2; in higher dimensions they are drawn after a volume-preserving whitening (det A = 1), and the volume is evaluated in log space for numerical stability.

Differentiation. Gradients flow through the boundary radii by implicit differentiation of Sθ(x, ρu) = τ, which gives ∇θ ρ = −∇θ S / (uᵀ ∇ᵣ S) at the boundary point, so no gradient passes through the root solver. The spline volumes are differentiated in closed form.

We thank you again for reviewing our work. Please let us know if we misunderstood any of your questions; we are happy to provide any further clarification.

References. [1] Braun et al., Minimum Volume Conformal Sets for Multivariate Regression, JASA 2026. [2] Yeh et al., End-to-End Conformal Calibration for Optimization Under Uncertainty, TMLR 2025.

Reviewer_mt3ZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CoCoNet proposes a framework for learning input-dependent convex conformity scores in multivariate regression. It introduces two parameterizations: a spline-based model producing rotated hyperrectangles and an ICNN-based model using a Bregman divergence to define flexible convex regions. The scores are trained to minimize prediction-region volume under soft coverage constraints, with the ICNN volume estimated through radial quasi-Monte Carlo integration and implicit differentiation. Standard split-conformal calibration ensures marginal coverage. Experiments on synthetic data and 30 real-world multivariate regression tasks show substantial improvements in region efficiency for the ICNN variant.

Critical strengths
  • The real strength of the paper is to combine a volume-minimization objective with a geometry whose volume can be computed differentiably (radial QMC integration with implicit differentiation of the boundary radii, App. C.3), while being, within the convex class, strictly more expressive than the norm-based families of Braun et al. (2026): the ICNN sublevel sets allow for substantially more flexible, asymmetric and input-dependent convex geometries than the norm-based families of Braun et al. The synthetic experiments (Fig. 2) make this geometric expressivity advantage visually convincing.
  • The empirical volume results on real datasets are strong: on 30 multivariate regression tasks, CoCoNet$_{\text{ICNN}}$ attains the smallest dimension-normalized volume on 25 tasks and never ranks below third.
Critical weaknesses
  • The methodological novelty is a combination of existing components: the volume-minimization objective of Braun et al. (2026) and the (P)ICNN parameterization of convex conformal scores already used by Yeh et al. (2025), which the paper cites only as a decision-focused method without discussing the direct methodological overlap in learning conditional convex conformity scores. Moreover, learning the uncertainty geometry through an input-convex network is also the core of Kondratyev et al. (2026), compared in App. E but never discussed as a methodologically adjacent construction, and of Conrad et al. (2026), which is not cited.
  • Given the empirical results, the lack of novelty in the ingredients would not by itself be a problem. What is critical is that, at the current stage of the conformal literature, it is not acceptable to report no measure of conditional coverage at all — at a minimum worst-slab coverage (Cauchois et al., 2021), and optionally the excess-risk diagnostics of Braun et al. (2026). This matters especially here: volume minimization is known to degrade conditional coverage, as Braun et al. (2026) themselves acknowledge in their discussion, and CoCoNet optimizes volume with a more expressive, input-dependent geometry than MVCS, making an empirical assessment of this potential trade-off very important. The paper reports only marginal coverage and defers conditional coverage to future work.

The following weaknesses are not blocking, but I think they are worth raising.

  • CoCoNet optimizes the region volume at a fixed training threshold τtr=1\tau_{\mathrm{tr}} = 1, and then builds the final regions with the conformal quantile τ^α\hat\tau_\alpha. In general, however, Vθ(x;1)≠Vθ(x;τ^α),V_\theta(x; 1) \neq V_\theta(x; \hat\tau_\alpha),.
  • There is no theoretical analysis of the efficiency or optimality of the approach. The stated results only show that admissible scores yield convex regions and that split calibration is valid; nothing connects the training objective to the population problem it approximates. The representation result says the class is rich enough, not that the procedure finds a good element of it.
  • Some relevant baselines are missing. The experiments do not compare against OT-CP (Thurin et al., 2025), CONTRA (Fang et al., 2025) or GCQR (Conrad et al., 2026). These comparisons would help assess the efficiency trade-off associated with imposing convexity.
  • Five seeds is a small number for the headline claim that CoCoNet$_{\text{ICNN}}$ ranks first on 25 of 30 tasks. On several datasets the reported standard deviations overlap with the second-best method.
Questions for authors
  • Could the authors report worst-slab coverage (Cauchois et al., 2021) for all compared methods on the 30 real datasets, and, if possible, the excess-risk conditional-coverage diagnostics of Braun et al. (2025)? Could they also add baselines :OT-CP (Thurin et al., 2025), CONTRA (Fang et al., 2025), and GCQR (Conrad et al., 2026)?
  • The score is trained to minimize volume at τtr=1\tau_{\mathrm{tr}} = 1, while the final regions use τ^α\hat\tau_\alpha. Could the authors report the distribution of τ^α\hat\tau_\alpha across datasets and seeds, and the relative difference between Vθ(x;1)V_\theta(x;1) and Vθ(x;τ^α)V_\theta(x;\hat\tau_\alpha)? If τ^α\hat\tau_\alpha is often far from 1, would training at the conformal threshold change the results?
8 comments
AuthorsOpenReview ↗

Response to Reviewer mt3Z (1/3)

We appreciate your constructive feedback, and we are glad that you found the combination of a volume objective with a differentiable volume for a more expressive convex class, and the empirical results, to be the strengths of the paper. We provide point-by-point responses below (in three parts); all changes in the revised PDF are marked in blue.

Regarding comparison with Yeh et al. and MVCS

Thank you for these pointers. The main distinctions from these works are that CoCoNet is agnostic both to the downstream decision problem and to the point predictor. It minimizes the size of the prediction region directly, so it requires no prior knowledge of how the regions will be used and the same regions can serve any downstream problem; and it learns scores of the form S(x, y − F(x)) on top of any pretrained point predictor F. Moreover, CoCoNet is a framework rather than a single geometry: CoCoNet-ICNN provides smooth, flexible convex sets, while the spline variants produce efficient structured sets (rotated boxes and polytopes), so practitioners can choose the degree of flexibility and structure that suits their task. The revised introduction now includes a paragraph on the two closest works (MVCS and Yeh et al.), and the related work (Appendix A) discusses all of them explicitly:

Yeh et al. [1] also learn input-dependent convex sets with an input-convex network. CoCoNet differs in two ways:

  • Learning objective. CoCoNet directly minimizes prediction-region size, whereas Yeh et al. optimize the performance of one downstream decision problem. CoCoNet therefore applies when the downstream problem is not known in advance, and the same regions can be reused across problems.
  • Score parameterization. The CoCoNet-ICNN score is a Bregman divergence, so it is nonnegative and vanishes at zero residual by construction, and its softplus activations give smoothly curved convex regions. Yeh et al. use ReLU activations to keep the downstream robust problem tractable, which restricts their regions to polytopes. CoCoNet also covers polytopes, through the new CoCoNet-Spline-ℓ1 variant, which learns input-adaptive, asymmetric polytopes whose size is minimized directly (mean rank 2.97 in Table 2, second only to CoCoNet-ICNN).

MVCS [2] also minimizes region volume, but over norm-based families. CoCoNet differs in two ways:

  • Symmetry. Single-norm MVCS uses one ℓq norm whose scale and orientation adapt to the input through a learned matrix, so its regions are symmetric about the point prediction. Our spline framework learns asymmetric, input-adaptive norms: both the shape of the norm and its asymmetry about the point prediction change with the input.
  • Input adaptivity of flexible shapes. The multi-norm variant of MVCS allows asymmetric shapes but is used as one global shape for all inputs. CoCoNet learns input-adaptive asymmetric regions of increasing flexibility: rotated boxes (CoCoNet-Spline-ℓ∞), rotated polytopes (CoCoNet-Spline-ℓ1) and smooth convex shapes (CoCoNet-ICNN).

Kondratyev et al. [3] (PBS) learn a pullback geometry that is not convex in general; PBS is now compared on all 30 real tasks (Table 10) and on the new non-convex synthetic targets (Appendix E.5). Conrad et al. [4] is now cited and discussed in Appendix A, and its GCQR and GRCP methods are included as baselines (Part 2).

Regarding random seeds

Thank you for this comment. All real-data experiments of the main comparison (Table 2 and Appendix Tables 8–10) are updated to ten random seeds, with the same protocol for every method. Mean rank over the 30 tasks among the eight methods of Table 2 (± standard deviation over tasks):

MethodMean rank ↓Tasks ranked first
CoCoNet-ICNN1.50 ± 0.9623
CoCoNet-Spline-ℓ1 (new)2.97 ± 1.431
MVCS3.47 ± 1.542
CoCoNet-Spline-ℓ∞4.30 ± 1.592
Local Ellipsoid4.57 ± 1.450
Global Ellipsoid5.00 ± 1.912
Abs. Residual7.10 ± 0.940
Empirical Copula7.10 ± 1.110

With ten seeds, CoCoNet-ICNN ranks first on 23 of the 30 tasks, is among the top three on 28 and is never below fourth. Table 2 reports the ten-seed standard deviations, so overlapping cells are visible. Since Table 2 reports the dimension-normalized size (mean volume)^{1/p}, the differences in actual volume are larger: for example, a 12.5% smaller normalized size on QM8 (p = 4) corresponds to a 41% smaller volume.

References. [1] Yeh et al., End-to-End Conformal Calibration for Optimization Under Uncertainty, TMLR 2025. [2] Braun et al., Minimum Volume Conformal Sets for Multivariate Regression, JASA 2026. [3] Kondratyev et al., Neural Optimal Transport Meets Multivariate Conformal Prediction, arXiv 2025. [4] Conrad et al., Geometric Conformal Prediction with Spatial Ranks and Multivariate Quantiles, ICML 2026.

AuthorsOpenReview ↗

Response to Reviewer mt3Z (2/3)

Regarding additional baselines

Thank you for this suggestion. OT-CP and OT-CP+ [1], CONTRA [2] and GCQR and GRCP [3] are now included, together with JAPAN [4], PBS [5] and free-q MVCS, on the same splits and the same frozen point predictor (Section 5; Appendix E.2, Table 10; Figure 4). Mean rank over the 30 tasks among the eleven methods of Table 10 (ten seeds; ± standard deviation over tasks), from best to worst:

MethodConvexMean rank ↓Tasks ranked first
CoCoNet-ICNNyes1.97 ± 1.3015
JAPANno2.73 ± 1.618
CoCoNet-Spline-ℓ1yes3.57 ± 1.581
MVCS (free q)no4.20 ± 1.901
CoCoNet-Spline-ℓ∞yes4.73 ± 1.390
CONTRAno6.40 ± 1.700
PBSno6.67 ± 1.641
OT-CP+no7.27 ± 3.054
OT-CPno8.82 ± 1.04 (28 tasks)0
GRCPno9.45 ± 1.75 (20 tasks)0
GCQRno9.50 ± 1.50 (26 tasks)0

CoCoNet-ICNN has the lowest mean rank and ranks first on 15 of the 30 tasks, followed by JAPAN (first on 8). So, while CoCoNet-ICNN offers the additional benefit of convex prediction regions, it remains the most efficient method overall, including against non-convex baselines.

Q1. Regarding worst-slab coverage

Thank you for this important suggestion. We now report the worst-slab coverage (WSC) [6] of all 15 methods on the 30 real tasks (slab mass 0.1, 1000 random directions, slab searched on 25% of the test split and evaluated on the remaining 75% so that the estimate is unbiased; Appendix E.3, Tables 11–12). Every method attains the nominal 90% marginal coverage. Summary, sorted by mean size rank:

MethodConvexSize rank ↓WSC (%) ↑Tasks with WSC < 85%
CoCoNet-ICNNyes2.2082.2 ± 11.715
JAPANno3.1383.1 ± 8.816
CoCoNet-Spline-ℓ1yes4.1077.4 ± 18.518
MVCSyes4.6778.4 ± 12.818
CoCoNet-Spline-ℓ∞yes5.7081.4 ± 11.116
Local Ellipsoidyes6.1785.0 ± 7.410
Global Ellipsoidyes6.9777.7 ± 12.420
CONTRAno8.0386.9 ± 5.77
PBSno8.9781.6 ± 7.320
OT-CP+no9.7389.4 ± 3.62
Empirical Copulayes10.4785.4 ± 6.611
Abs. Residualyes11.0780.9 ± 10.617
OT-CPno12.0482.6 ± 9.814
GRCPno12.6080.3 ± 9.618
GCQRno13.0881.4 ± 8.419

We agree that this was an important point to check, since a more expressive, input-dependent geometry could in principle trade conditional coverage for smaller regions. Reassuringly, on this diagnostic CoCoNet-ICNN attains a higher WSC than MVCS (82.2% against 78.4%) and has fewer tasks below 85% (15 against 18). More broadly, there is a trade-off between overall set size and conditional coverage: the methods with the highest WSC, OT-CP+ (89.4%) and CONTRA (86.9%), reach it with much larger regions (mean size ranks 9.73 and 8.03). CoCoNet-ICNN sits at a favorable point of this trade-off, with the smallest regions and a WSC close to JAPAN (83.1%), the second most efficient method. The revised conclusion names conditional coverage as a limitation and a direction for future work.

References. [1] Thurin et al., Optimal Transport-based Conformal Prediction, ICML 2025. [2] Fang et al., CONTRA: Conformal Prediction Region via Normalizing Flow Transformation, ICLR 2025. [3] Conrad et al., Geometric Conformal Prediction with Spatial Ranks and Multivariate Quantiles, ICML 2026. [4] English and Lippert, JAPAN: Joint Adaptive Prediction Areas with Normalising Flow, ICLR 2026. [5] Kondratyev et al., Neural Optimal Transport Meets Multivariate Conformal Prediction, arXiv 2025. [6] Cauchois et al., Knowing What You Know: Valid and Validated Confidence Sets in Multiclass and Multilabel Prediction, JMLR 2021.

AuthorsOpenReview ↗

Response to Reviewer mt3Z (3/3)

Q2. Regarding the training threshold τ = 1 and the conformal threshold q̂

Thank you for this question. The revised Section 3 now states explicitly that the training threshold is fixed at τ = 1 to simplify the optimization: the coverage terms steer the unit sublevel set towards coverage 1 − α, and the precise threshold is then set by calibration. The new Appendix E.4 (Table 13) reports q̂ for every task: over all tasks and seeds, the median is 1.68 (interquartile range 1.27–2.38) for CoCoNet-ICNN, 1.80 (1.34–2.58) for CoCoNet-Spline-ℓ∞ and 1.98 (1.54–4.38) for CoCoNet-Spline-ℓ1. Coverage is guaranteed by the calibration step whatever the value of q̂, and every size reported in the paper is that of the calibrated region, for every method.

We do not train at the conformal threshold for three reasons. First, q̂ is a quantile of the trained score on held-out calibration data, so it does not exist while the score is trained. Second, a jointly learned threshold would be redundant: for an admissible score S, the pairs (S, τ) and (S/τ, 1) define the same region, up to the regularization constants of each parametrization, so fixing τ = 1 removes this redundancy. Third, a fixed threshold keeps training numerically stable: with a learnable threshold, the scale of the score and the threshold could drift together during optimization, changing the scale of the boundary radii and of the volume objective computed from them.

Regarding your concern that the volume minimized at τ = 1 differs from the volume of the final region at q̂: training does not leave this gap to calibration alone. The coverage terms of the objective push the unit sublevel set towards coverage 1 − α, which keeps the size objective from shrinking the set far below the target coverage; calibration then adjusts its level. The new loss-term ablation (Appendix E.7, Table 20) shows that this matters for the final, calibrated size: removing the instance-level miss term enlarges the calibrated region on nine of the ten datasets (mean rank 2.80, against 1.20 for the full objective, which is best on eight of the ten).

We thank you again for reviewing our work. Please let us know if we misunderstood any of your questions; we are happy to provide any further clarification.

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

The paper studies the design of conformity scores for multivariate regression. Conformal prediction gives distribution-free marginal coverage guarantees, but the size and shape of the prediction regions depend heavily on the choice of score. The authors propose CoCoNet, which learns an input-dependent score that is convex in the residual, so that its sublevel sets are convex regions containing the point prediction. The score is trained on a held-out split with a loss that explicitly minimises region volume (plus soft coverage penalties), and is then calibrated with split conformal prediction, preserving finite-sample coverage. Two parameterisations are proposed: a piecewise-linear spline score (rotated asymmetric boxes) and an ICNN-based score (general convex shapes). The authors also show that any compact convex region containing the prediction can be represented as a sublevel set of such a score. Validation consists of synthetic 2D experiments with known optimal regions and a benchmark of 30 real datasets (output dimension 2–16) against convex baselines (coordinatewise, copula, global/local ellipsoids, MVCS), with additional comparisons to non-convex methods in the appendix. CoCoNet-ICNN achieves the smallest region size on 25 of 30 tasks while maintaining nominal coverage.

Critical strengths

The method is evaluated on 30 real-world multivariate regression tasks spanning output dimensions from 2 to 16, plus synthetic settings with known optimal regions, against a broad set of convex baselines (coordinatewise, copula, global/local ellipsoids, MVCS) and, in the appendix, non-convex alternatives (free-q and multi-norm MVCS, PBS). All methods share the same base predictor and data splits, results are averaged over five seeds, and coverage is reported alongside size. This is consequential because the paper's main claim is empirical (more efficient regions under valid coverage), and the breadth of tasks and baselines makes that claim more credible than a small set of hand-picked datasets would.

Critical weaknesses
  1. Representing conformal prediction regions as sublevel sets of a learned score is not new. MVCS (Braun et al.) already learns input-adaptive conformal sets by minimising volume (see also the concurrent Super-Level-Set Regression of Braun, Jordan & Bach, 2026). The theoretical results are also standard: Proposition 2 and Proposition 3 are textbook facts about sublevel sets of a convex function.
  2. Convexity is not justified. The paper motivates convex regions by interpretability, regularisation and tractability in downstream robust optimisation, but none of these is evaluated. Convexity is a strong restriction. Highest-density regions are convex at every level only for quasiconcave densities; even distributions with a single connected high-density region (e.g. banana-shaped) generally have non-convex regions, and for multimodal responses any convex set must also cover the low-density gaps between modes. Requiring the point prediction to lie inside the region makes this worse, since a mean-regression prediction can fall in such a gap. The synthetic study only uses convex targets (ellipse, rectangle, triangle, trapezoid), so it never tests the regime where the restriction is costly.
Questions for authors

Clarification requests

  1. The paper motivates convex regions by interpretability, regularisation and tractability in downstream robust optimisation, but none of these is evaluated. Could you provide concrete evidence that convexity is necessary or beneficial ?

  2. The radial estimator in eq. (56) assumes each ray from F(x) crosses the region boundary exactly once, which holds for star-shaped sets but not in general. How were the volumes of PBS and the free-q MVCS variants in Table 10 computed ?

  3. All synthetic targets are convex (ellipse, rectangle, triangle, trapezoid). Could you add synthetic settings whose optimal regions are non-convex (e.g. banana-shaped, ring-shaped, or bimodal), report IoU and V/V* there, and compare against more flexible methods (e.g. PBS, CONTRA, JAPAN, or density-based scores) ?

  4. Since the score is highly input-adaptive and trained to minimise volume under a marginal coverage constraint, smaller regions could come at the cost of conditional coverage. Could you report a conditional coverage diagnostic for CoCoNet and the baselines?

  5. The ablation suggests most of the gain comes from the contextual encoder. How does a global (input-independent) CoCoNet_ICNN compare to Global Ellipsoid on the full benchmark? Conversely, would MVCS or Local Ellipsoid equipped with the same encoder h[x, F(x)] close the gap?

5 comments
AuthorsOpenReview ↗

Response to Reviewer orPx (1/3)

We appreciate your careful feedback and your recognition of the breadth of the benchmark. We provide point-by-point responses below (in three parts); all changes in the revised PDF are marked in blue.

Regarding novelty and comparison with MVCS

We agree that CoCoNet and MVCS share a goal: both learn conformity scores that produce efficient prediction regions. However, there are a few key distinctions:

  • Single-norm MVCS learns an input-dependent transformation with a single norm exponent qq shared across inputs, so each region is symmetric around its center. This symmetry limits its ability to represent asymmetric uncertainty, even when the underlying region is convex.
  • Multi-norm MVCS enables asymmetric regions by combining several norms on different parts of the residual space [1, Prop. 2.6], but it is used as one global shape for all inputs: the authors report that "extending the multi-norm MVCS framework to locally adaptive transformations in regression often leads to overfitting" [1, Sec. 3].
  • CoCoNet learns input-dependent convex scores directly in residual space. The revised paper adds CoCoNet-Spline-ℓ1\ell_1, which uses ℓ1\ell_1-norm aggregation to produce rotated asymmetric polytopes. Together with CoCoNet-Spline-ℓ∞\ell_\infty, which produces rotated asymmetric boxes, and CoCoNet-ICNN, which learns smooth convex regions of flexible shape, the new variant shows that CoCoNet is a single framework for learning input-adaptive, asymmetric convex regions with different degrees of geometric flexibility.

We have also revised the introduction and the extended literature review (Appendix A) to clarify these points and the comparison with other relevant approaches.

Q1. Regarding convexity

Thank you for your comment. We have revised the convexity paragraph of the introduction to make this motivation explicit. Convexity is desirable for three reasons:

  • Geometric regularization. Convexity rules out the fragmented or disconnected regions that flexible non-convex scores can produce in data-sparse areas, while the region can still adapt its orientation and scale to heterogeneous uncertainty geometries.
  • Tractable downstream decisions. Membership testing, projection and support-function evaluation are efficient over convex regions, and the "must hold for every point in the set" constraints of robust optimization, planning and control admit tractable reformulations via duality [2, 3]. Convex CP regions can therefore be used directly as uncertainty sets in these frameworks while retaining the coverage guarantee of CP [4, 5].
  • Size is only saved where it matters. For constraints that are convex in the outcome, the worst case over any set equals the worst case over its convex hull [2, 3]. A non-convex region that is smaller in volume therefore gives the same robust decision as its convex hull, so its saving in volume does not carry over to the decision.

Nevertheless, we added seven non-convex baselines and show that the proposed approach still provides the best overall rank on the real datasets (Q3.1 and Q3.2, Part 2).

Q2. Regarding volume estimation for PBS and free-qq MVCS

Thank you for your comment. As you point out, PBS regions need not be star-shaped, so we use the volume estimator of PBS itself, which relies on the change of variables through its pullback map [6]. As a check, we also traced every boundary crossing along each ray on 26,030 test points (p=2,4,8p = 2, 4, 8); the two volumes agree to within 0.07% in dimension-normalized size (0.5% in volume).

Single-norm MVCS sizes (convex and free-qq) use the exact closed-form ℓq\ell_q-ball volume; for the multi-norm variants, which have no closed form, we use the same radial estimator as for CoCoNet-ICNN. It only needs each ray to cross the boundary once, which holds for any region that is star-shaped about the ray origin; an ℓq\ell_q ball is star-shaped about its center for every q>0q > 0. In our experiments, the free-exponent MVCS selected q≥1.03q \ge 1.03 in 299 of its 300 cells (ten seeds), which results in convex regions; the exception is Online Retail II with seed 9 (q=0.93q = 0.93). The multi-norm variant selected q<1q < 1 for at least one of its norms in 16 of its 300 cells; its regions are still star-shaped, so the estimator remains exact. We clarified this point in Appendix C.3 and in the caption of Table 18.

References. [1] Braun et al., Minimum Volume Conformal Sets for Multivariate Regression, JASA 2026. [2] Ben-Tal, El Ghaoui and Nemirovski, Robust Optimization, 2009. [3] Bertsimas, Brown and Caramanis, Theory and Applications of Robust Optimization, SIAM Review 2011. [4] Johnstone and Cox, Conformal Uncertainty Sets for Robust Optimization, COPA 2021. [5] Tumu et al., Multi-Modal Conformal Prediction Regions by Optimizing Convex Shape Templates, L4DC 2024. [6] Kondratyev et al., Neural Optimal Transport Meets Multivariate Conformal Prediction, arXiv 2025.

AuthorsOpenReview ↗

Response to Reviewer orPx (2/3)

Thank you for this suggestion. While the main goal of CoCoNet is to provide efficient convex prediction regions, we agree that it is informative to measure what the convexity constraint costs relative to methods that do not impose it. We therefore added non-convex baselines and non-convex synthetic targets (Q3.1 and Q3.2).

Q3.1. Additional non-convex baselines

On real data, we added seven non-convex baselines: OT-CP and OT-CP+ [5], CONTRA [2], JAPAN [3], GCQR and GRCP [4], and PBS [1], next to free-qq MVCS. All use the same splits and the same frozen point predictor as CoCoNet. Each method is evaluated with the volume estimator of its own paper. The results are in Table 10 (Appendix E.2) and Figure 4 (Appendix E.1). Mean rank over the 30 tasks among the eleven methods of Table 10 (ten seeds; ±\pm standard deviation over tasks), from best to worst:

MethodConvexMean rank ↓Tasks ranked first
CoCoNet-ICNNyes1.97 ± 1.3015
JAPANno2.73 ± 1.618
CoCoNet-Spline-ℓ1\ell_1yes3.57 ± 1.581
MVCS (free qq)no4.20 ± 1.901
CoCoNet-Spline-ℓ∞\ell_\inftyyes4.73 ± 1.390
CONTRAno6.40 ± 1.700
PBSno6.67 ± 1.641
OT-CP+no7.27 ± 3.054
OT-CPno8.82 ± 1.04 (28 tasks)0
GRCPno9.45 ± 1.75 (20 tasks)0
GCQRno9.50 ± 1.50 (26 tasks)0

CoCoNet-ICNN has the lowest mean rank and ranks first on 15 of the 30 tasks, ahead of JAPAN (2.73, first on 8). On these real-data benchmarks, which were not selected for convexity. OT-CP+ regions are unbounded in at least one seed on all 30 tasks, OT-CP regions on 22 and GRCP regions on 3, so their sizes are lower bounds; OT-CP, GCQR and GRCP are not measurable on 2, 4 and 10 tasks (the region is too small for a reliable Monte Carlo estimate) and are ranked where measurable. Since the tables report the dimension-normalized size (mean volume)1/p^{1/p}, differences in actual volume are larger.

Q3.2. Additional synthetic instances

As you suggested, we added three non-convex targets (banana, ring and bimodal; Appendix E.5, Table 15, Figure 5), generated like the convex ones so that the minimum-area 90% region is known. We report IoU and V/V⋆V/V^\star for 13 methods, including PBS, CONTRA, JAPAN and OT-CP (five seeds; coverage 89.4–90.8% for every method). A selection (IoU ↑ / V/V⋆V/V^\star ↓):

MethodConvexBananaRingBimodal
CoCoNet-ICNNyes0.511 / 1.840.697 / 1.360.611 / 1.54
MVCSyes0.356 / 2.530.707 / 1.360.577 / 1.62
Global Ellipsoidyes0.352 / 2.560.708 / 1.350.551 / 1.69
JAPANno0.800 / 1.210.685 / 1.390.741 / 1.29
CONTRAno0.690 / 1.370.576 / 1.610.602 / 1.55
OT-CPno0.513 / 1.850.701 / 1.370.585 / 1.61
PBSno0.334 / 2.770.699 / 1.370.395 / 2.58

CoCoNet-ICNN is the only method, convex or not, whose region is among the three smallest of all 13 on every target: second on the bimodal target (behind JAPAN), joint second on the ring (matching MVCS at V/V⋆V/V^\star = 1.36, ahead of JAPAN and CONTRA) and third on the banana. It is also the most accurate convex method on the banana and bimodal targets, and on the banana its area is 27% smaller than that of MVCS (1.84 against 2.53). The density-based JAPAN and CONTRA recover the banana more closely (IoU 0.800 and 0.690, against 0.511), which reflects the price of convexity on strongly non-convex targets; the appendix now discusses this explicitly.

References. [1] Kondratyev et al., Neural Optimal Transport Meets Multivariate Conformal Prediction, arXiv 2025. [2] Fang et al., CONTRA: Conformal Prediction Region via Normalizing Flow Transformation, ICLR 2025. [3] English and Lippert, JAPAN: Joint Adaptive Prediction Areas with Normalising Flow, ICLR 2026. [4] Conrad et al., Geometric Conformal Prediction with Spatial Ranks and Multivariate Quantiles, ICML 2026. [5] Thurin et al., Optimal Transport-based Conformal Prediction, ICML 2025.

AuthorsOpenReview ↗

Response to Reviewer orPx (3/3)

Q4. On conditional coverage

Thank you for this suggestion. While CoCoNet does not provide conditional guarantees, we agree that conditional coverage is important to report, as both you and reviewer mt3Z pointed out. We now report the worst-slab coverage (WSC) [1] of all 15 methods on the 30 real tasks (slab mass 0.1, 1000 random directions, slab searched on 25% of the test split and evaluated on the remaining 75%; Appendix E.3, Tables 11–12). Every method attains the nominal 90% marginal coverage. Summary, sorted by mean size rank:

MethodConvexSize rank ↓WSC (%) ↑Tasks with WSC < 85%
CoCoNet-ICNNyes2.2082.2 ± 11.715
JAPANno3.1383.1 ± 8.816
CoCoNet-Spline-ℓ1yes4.1077.4 ± 18.518
MVCSyes4.6778.4 ± 12.818
CoCoNet-Spline-ℓ∞yes5.7081.4 ± 11.116
Local Ellipsoidyes6.1785.0 ± 7.410
Global Ellipsoidyes6.9777.7 ± 12.420
CONTRAno8.0386.9 ± 5.77
PBSno8.9781.6 ± 7.320
OT-CP+no9.7389.4 ± 3.62
Empirical Copulayes10.4785.4 ± 6.611
Abs. Residualyes11.0780.9 ± 10.617
OT-CPno12.0482.6 ± 9.814
GRCPno12.6080.3 ± 9.618
GCQRno13.0881.4 ± 8.419

There is a trade-off between overall set size and conditional coverage: the methods with the highest WSC, OT-CP+ (89.4%), CONTRA (86.9%) and Empirical Copula (85.4%), reach it with much larger regions (mean size ranks 9.73, 8.03, 10.47). CoCoNet-ICNN sits at a favorable point of this trade-off: it has the smallest regions (mean size rank 2.20) with a mean WSC of 82.2%, above MVCS (78.4%) and close to JAPAN (83.1%), the second most efficient method. The revised conclusion names conditional coverage as a limitation and a direction for future work.

Q5. Ablation and the role of the encoder

Thank you for this question. Input dependence is a central part of CoCoNet: the encoder h[x,F(x)]h[x, F(x)] is trained end to end with the convex score and its volume objective, and the ablation (Appendix E.7, Table 19) confirms that it matters (mean rank 1.60 for the full CoCoNet-ICNN against 6.10 for its input-independent version). The efficiency of CoCoNet, however, does not come from the encoder alone:

  • Equipping the baselines with the same encoder. Because the encoder is learned end to end with CoCoNet's score and volume objective, it is not a component that can be plugged into MVCS or the Local Ellipsoid; doing so would require a new training objective for each method. Both baselines also already adapt to the input in their own way: the Local Ellipsoid estimates its covariance from nearest neighbors in input space, and MVCS learns its input-dependent matrix with its own neural network on xx. To keep the comparison faithful to the published methods, we kept every baseline as its authors defined it.
  • The learned geometry. The advantage of CoCoNet grows as the underlying geometry becomes more complex. In the synthetic study (Table 1), where the residual distribution is the same for every input, CoCoNet-ICNN is close to MVCS and the Global Ellipsoid on the ellipsoidal target (IoU 0.94 against 0.97), but clearly the best on the triangle (0.85 against 0.57) and the trapezoid (0.83 against 0.68 and 0.67). These gains come from learning a more flexible convex shape rather than from input dependence.

We thank you again for reviewing our work. Please let us know if we misunderstood any of your questions; we are happy to provide any further clarification.

References. [1] Cauchois et al., Knowing What You Know: Valid and Validated Confidence Sets in Multiclass and Multilabel Prediction, JMLR 2021.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We thank all reviewers for their detailed feedback, which helped us improve the paper. Below is a summary of the main changes made after the rebuttal; all changes are highlighted in blue in the revised PDF.

  • New baselines. We added seven non-convex baselines (OT-CP, OT-CP+, CONTRA, JAPAN, GCQR, GRCP, PBS) and the free-qq and learned-centre MVCS variants on all 30 real tasks. CoCoNet-ICNN remains the best convex method and outperforms all seven non-convex baselines. The only method with smaller regions is the non-convex MVCS with free qq and a learned centre, which learns the centre of its regions, whereas CoCoNet builds them around the given point prediction. While CoCoNet-ICNN ranks slightly behind it (3.23 vs. 3.10), it is more stable across tasks (rank standard deviation 2.30 vs. 3.22) and has better worst-slab conditional coverage (81.4% vs. 78.7%).

  • Downstream decision task. We added a robust decision task as evidence of the value of efficient convex uncertainty regions. At the same coverage, CoCoNet-ICNN needs the smallest reservation and has the lowest over-procurement, and the convex hulls of non-convex regions are all larger than the CoCoNet-ICNN region.

  • Stronger baselines. We tuned the Local Ellipsoid, fixed the GCQR training and GRCP bandwidth selection, and used whitened residuals for both. The main conclusions are unchanged.

  • Conditional coverage. We added worst-slab conditional coverage for all methods. CoCoNet-ICNN gives the smallest regions, with conditional coverage above MVCS and close to the best method.

  • Ablations. We added ablations of the loss terms and of the input-dependent score, and a paragraph separating the contribution of the framework from that of the score family (Appendix E.8). Every component contributes, and efficiency grows with the flexibility of the score family.

  • Reproducibility. We released the code and configuration of CoCoNet and of every baseline, and the main results now use ten seeds. Six of the baselines (CONTRA, PBS, the Global and Local Ellipsoid, the Empirical Copula and LW-SC) have no public code; we re-implemented them as closely as possible to the original papers and released their implementations.

  • New variant. We added CoCoNet-Spline-ℓ1\ell_1 (rotated asymmetric polytopes), which improves on the box-shaped spline variant and is the second-best convex method after CoCoNet-ICNN.

  • Presentation. We revised the abstract, introduction and Figure 1, moved the positioning against MVCS and Yeh et al. into the main text, clarified the training threshold and volume estimation, and redesigned the main results table to make it easier to read.

We hope these changes address the reviewers' concerns. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.