acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Conformal Prediction via Learned Residual-Law Retrieval

Abstract

Conformal prediction provides finite-sample marginal coverage, but constructing efficient prediction sets that adapt to query-dependent uncertainty remains challenging. Current approaches generally obtain local adaptivity either by localizing calibration or by learning a directly parameterized conditional uncertainty model. We introduce Factorized Conformal Prediction (FacCP), which factorizes learned local residual-law retrieval from global calibration. The retrieved law adapts prediction geometry to each query, while conformal calibration remains a single global scalar estimation problem. We show that accurate local geometry induces approximately pivotal scores, making one global quantile approximately appropriate across queries while preserving finite-sample marginal validity. We bound conditional coverage error through geometry approximation and global calibration error, and decompose geometry error into representation, localization, estimation, and geometry-class contributions. FacCP learns an uncertainty-relevant representation with an energy-score objective targeting the full conditional residual law, retrieves this law locally, and instantiates query-specific geometries; the same representation supports multiple set constructors without retraining. Controlled oracle experiments validate the predicted chain from residual-law recovery to geometry, pivotality, and conditional coverage, and show that even oracle local calibration can remain geometrically inefficient. Across high-dimensional robotic reachability tasks, FacCP maintains calibrated coverage and stable set efficiency across wide variations in model capacity and training data, while direct covariance prediction becomes more efficient only in higher-data, higher-capacity regimes.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conformal Prediction, Uncertainty Quantification, Distributional Learning, Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_FmkNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FacCP, which separates local uncertainty estimation from global conformal calibration. Given a fixed predictor and a task-specific residual map T(X,Y)T(X,Y), an input representation hθh_\theta is trained using a leave-one-out energy-score objective. Kernel weights in the learned representation space are then used to estimate the conditional residual law PxRP_x^R as a weighted empirical distribution over a reference set. A geometry constructor Γ\Gamma (ellipsoid, hyperbox, KDE, or zonotope) maps the estimated law to nested residual sets, which induce a scalar nonconformity score calibrated using a single global split-conformal quantile. The theoretical analysis characterizes uncertainty-sufficient representations, studies the localization penalty induced by nuisance dimensions, and bounds conditional-coverage error in terms of score non-pivotality, which is itself controlled under a uniform geometry-approximation assumption. Experiments include a synthetic bivariate Gaussian setting, planar and 3D quadcopter reachability with trajectory-level conformal calibration, and autonomous-driving trajectory prediction on Argoverse 2, with capacity and data sweeps against resCONTRA and direct covariance prediction.

Critical strengths
  • Learning the retrieval weights from an input representation trained with an energy-score objective, so that proximity reflects similarity of conditional residual laws rather than raw input proximity, is an interesting idea that could be valuable in high dimension.
  • The experiments are indeed carried out on high-dimensional problems, although only one dataset is non-synthetic.
  • One objection to this way of localizing the uncertainty score (here, the multivariate residuals) is that, unlike in Han et al. (Split Localized Conformal Prediction) or Plassier et al. (Rectifying conformity scores for better conditional coverage), the resulting score is not guaranteed pivotal. Under continuity and strict monotonicity of the score CDF at q1−αq_{1−\alpha}, those constructions allow the conditional coverage of the oracle score to be read off directly, so that conditional coverage error is quantified by the learning error of the score with respect to its oracle. Here this is no longer possible: one must additionally account for the pivotality gap of the oracle itself, analyzed through Δpiv\Delta_{\text{piv}}, which the authors attempt to do.
Critical weaknesses
  • However, Δpiv\Delta_{\text{piv}} is then left uncontrolled: Thm. 4 assumes the geometry-approximation property (Assumption 3) that the energy-score objective should deliver, and Thm. 11 reduces to a triangle inequality plus DKW once Δpiv\Delta_{\text{piv}}, is given. Nothing connects the design choice — the energy-score objective — to εgeom\varepsilon_{\text{geom}} or Δpiv\Delta_{\text{piv}}, even for the ellipsoidal constructor actually used, where equal-covariance laws with different tails already break pivotality. What is proved is standard; what is specific to FacCP is assumed. The analysis therefore names the right quantity but does not analyze it.
  • The fact that the multivariate residuals are treated as a given object, produced by a black-box model, is not made sufficiently clear when the authors state that they do not learn a score. The residual T(X,Y) is itself a fixed vector-valued conformity score; the paper’s framing of “estimating the residual law rather than learning a score” conflates T with the induced scalar score S^\hat{S}. What FacCP actually does is learn an adaptive scalarization of a fixed multivariate score, which is precisely the setting addressed by recent multivariate conformal methods (Thurin et al., 2025; Klein et al., 2026; Kondratyev et al., 2026, Conrad et al., 2026) and should be positioned as such (even if they cite Thurin et al.).
  • The experiments do not isolate the advantage of the proposed representation over alternative modeling choices: no ablation replaces the energy-score objective by an alternative representation objective while keeping retrieval, geometry construction and calibration fixed, and on Argoverse the pretrained encoder alone (without energy fine-tuning) is never evaluated, so the gain attributable to the method’s central component cannot be read off the results. Moreover, the comparisons are mainly against resCONTRA, direct covariance prediction and localized-quantile baselines; NLE and MVCS appear only on a single-mode. Since the method is an adaptive scalarization of a fixed multivariate residual score, the natural comparators are the other methods that learn or construct a residual geometry from such a score, e.g. OT-CP (Thurin et al., 2025), MultiDimSPCI (Xu et al.,, 2024), CopulaCP (Messoudi et al., 2021), CVQR (Kondratyev et al., 2026), GCQR/GRCP (Conrad et al., 2026); none is evaluated. Several of these can be instantiated as the constructor Γ\Gamma on the retrieved law, which would turn this missing comparison into an ablation of the authors’ own framework.
Questions for authors
  • Across the capacity and data sweeps, could you investigate the empirical relationship between the held-out energy score of the retrieved residual law, the pivotality defect Δpiv\Delta_{\mathrm{piv}}, and prediction-set efficiency at comparable coverage? In particular, do improvements in residual-law estimation translate into better pivotality and tighter prediction sets?

  • For the ellipsoidal constructor used in practice, under which condition on PxRP^R_x does an energy-distance error on P^xR\hat{P}^R_x bound εgeom\varepsilon_{\mathrm{geom}} in Assumption 3? Can you state and prove such a bound for at least one constructor?

  • Since Γ\Gamma is arbitrary, can other scalarizations (like OT-CP) be instantiated as Γ\Gamma on the retrieved law, and how do they compare with the ellipsoidal constructor?

  • Clarification, not score-relevant: please unify the notation ZZ / UU / SS for the localization variable, state explicitly that ZZ enters only through the weights wj(x)w_j(x). More generally, the notation is heavy and partly redundant; lightening it would make the paper easier to follow.

5 comments
AuthorsOpenReview ↗

Response to Reviewer FmkN - Part 1

We thank the reviewer for the detailed feedback. We have revised the paper to clarify its central theoretical contribution, added a new constructor-specific pivotality result, and expanded the experimental comparisons. For this reviewer, we provide Highlights\_FmkN.pdf with relevant revisions highlighted for a targeted re-read.

1. Theoretical novelty, motivation, and pivotality

  • Novelty. (Weakness 1) We disagree that our theoretical contribution is standard. Our central result establishes a connection between local geometry approximation and score pivotality. Theorem 3 proves that approximating query-dependent oracle geometries with invariant conditional probability mass yields approximately pivotal scores, with Δpiv≤2ϵgeom\Delta_{\mathrm{piv}}\leq 2\epsilon_{\mathrm{geom}}. Unlike prior analyses based on score-distribution correction (Laplante, 2026; Plassier et al. 2025a) or conditional-law estimation error (Plassier et al., 2025b), Theorem 3 directly bounds score non-pivotality in terms of the conditional probability mass where the learned and true geometric sublevel sets disagree. Corollary 7 and Theorem 8 translate this into local coverage guarantees under global calibration, while Corollary 9 bounds coverage loss under covariate shift without score reweighting. This is important because correcting scalar thresholds cannot recover misspecified geometry. Figure 2 shows that even oracle local quantile calibration of a fixed geometry produces sets with 1.99×1.99\times oracle area.

  • Energy-score retrieval is a practical implementation of the theory. (Weakness 1) FacCP provides a practical realization of Theorem 3 by constructing query-dependent residual geometries from energy-score-trained residual law retrieval and aiming to make the resulting nonconformity scores approximately pivotal, doing so in a principled, scalable, and efficient manner. Theorem 3 applies to any nested geometry satisfying Assumption 2, regardless of how the geometries are constructed. Rather than directly estimating high-dimensional geometry parameters, FacCP learns a compact uncertainty-relevant representation using a strictly proper energy score that targets the full conditional residual law. Proposition 4 characterizes uncertainty sufficiency through excess energy-score risk, while Proposition 12 establishes the statistical benefit of suppressing uncertainty-irrelevant dimensions. Figure 3 and Appendix C.3 further support this design choice, demonstrating data and parameter efficiency against direct covariance and flow-based models, especially with limited training data or capacity.

  • Controlling pivotality. (Weakness 1, Question 2) We added Theorem 14, which bounds pivotality for the ellipsoidal constructor using energy-distance retrieval error. The bound separates non-pivotality inherent to the geometry class from error due to imperfect retrieval and covariance regularization. Crucially, it accounts for the reviewer's example of equal-covariance distributions with different tails. Even perfect residual-law estimation need not yield pivotal ellipsoidal scores when the chosen geometry cannot represent differences in conditional tail behavior. Empirically, the full-CDF pivotality defect decreases from 0.361 for global geometry to 0.117 with FacCP (Figure 2, Table 2).

2. Adaptive scalarization and requested multivariate conformal comparisons

  • Score definition. (Weakness 2) We agree that T(X,Y)T(X,Y) is a fixed vector-valued residual score and that FacCP learns its adaptive scalarization. We now explicitly distinguish this fixed residual map from the induced scalar nonconformity score S^(X,Y)\widehat S(X,Y), which measures the inflation required for the residual to enter the query-dependent geometry. See Sec. 3, 4.1, and 4.3 for the specific discussion.

  • Positioning relative to prior methods. (Weakness 2) We clarified FacCP's relationship to recent multivariate conformal methods (Thurin et al., 2025; Klein et al., 2026; Kondratyev et al., 2026; Conrad et al., 2026). Unlike methods focused on specific score geometries, FacCP establishes a general geometry-to-pivotality connection and learns residual-law retrieval independently of the geometry constructor (Section 2, Appendix B). Our choice of local residual-law modeling is motivated by its scalability, demonstrated empirically against direct geometry models (Figure 3). Theorem 3 applies to any nested geometry satisfying Assumption 2, and Appendix E.1 shows that learned retrieval also improves alternative constructors such as GRCP and CopulaCP.

AuthorsOpenReview ↗

Response to Reviewer FmkN - Part 2

  • Thurin et al.'s OT-CP+. (Weakness 3, Question 3) We added OT-CP+ comparisons. On our planar quadcopter benchmark, the 100-step prediction problem has a 600-dimensional joint residual space. The evaluated finite-reference OT-CP+ construction produces unbounded prediction regions due to rank deficiency in its transport representation. A per-timestep 6D adaptation reduces this dimensionality problem but still produces unbounded regions and requires 374.83 ms/query, compared with 1.74 ms/query for FacCP (Table 20).

  • Kondratyev et al.'s CVQR. (Weakness 3, Question 3) FacCP achieves comparable normalized log-volumes while requiring 152×152\times less geometry-construction time and 382×382\times less membership-evaluation time while maintaining empirical coverage (Table 21).

  • Conrad et al.'s GRCP. (Weakness 3, Question 3) Replacing GRCP's input-space retrieval with our learned retrieval improves normalized log-volume from −28.254-28.254 to −29.854-29.854 while retaining coverage (Table 23).

  • Messoudi et al.'s CopulaCP (2021). (Weakness 3, Question 3) FacCP produces substantially smaller sets than CopulaCP (log-volume −34.899-34.899 vs. −22.612-22.612). Combining FacCP geometry with CopulaCP calibration further reduces log-volume to −35.116-35.116 at comparable coverage, with minimal runtime overhead (Table 23).

  • Xu et al.'s MultiDimSPCI (2024). (Weakness 3, Question 3) MultiDimSPCI targets non-exchangeable multivariate time series through sequential adaptive calibration, whereas our setting assumes exchangeable trajectory samples and learns query-dependent geometry.

3. Missing representation and objective ablations

We added controlled ablations holding downstream retrieval, geometry construction, and calibration fixed.

  • Pretrained representations. (Weakness 3) On Argoverse 2, energy-score-trained retrieval produces substantially tighter uncertainty tubes than retrieval using pretrained MAE or frozen MTR features, while maintaining comparable whole-trajectory coverage. Energy-trained representations reduce footprint by 14.15% at the 95% target, compared with −0.12%-0.12\% for pretrained MAE features and 3.74% for frozen MTR features. Corresponding coverages are 94.87%, 95.25%, and 95.20% (Table 11).

  • Alternative training objective. (Weakness 3) To show the importance of the energy score, we replace energy-score fine-tuning with a residual-neighbor KL objective. This alternative objective reduces the footprint improvement from 14.15% to 0.76%, using matched architectures, initialization, and training settings (Table 12).

  • Quadcopter ablations. (Weakness 3) Retrieval using energy-trained representations achieves normalized log-volume −34.899-34.899 at the 95% target, compared with −28.903-28.903 for a covariate-reconstruction autoencoder and −27.466-27.466 for a residual-prediction autoencoder, with coverage near nominal (Table 6).

  • NLE and MVCS. (Weakness 3) Both methods fit a single joint ellipsoid, so we evaluate them in the single-mode setting. Extending them to six diverging trajectories would include large regions between paths. At the 95%95\% target, FacCP achieves a footprint of 1,882.41{,}882.4 m$^2$s versus 47,625.847{,}625.8 for NLE and 168,608.7168{,}608.7 for MVCS, all at comparable coverage (Table 20).

4. Does better held-out energy score correlate with pivotality and efficiency? (Question 1)

  • Theory and empirical validation. Theorem 14 bounds the estimation component of the pivotality defect using energy-distance retrieval error. Since the data and capacity sweeps showed little energy-score variation, we evaluated ten energy fine-tuning checkpoints after one epoch of MAE pretraining, keeping retrieval temperature, geometry, and calibration fixed (Appendix D.10).

  • Pivotality. Lower held-out energy scores correlate strongly with smaller pivotality defects across 14 scene bins, with Spearman correlations of ρ=0.915\rho=0.915 at 97% and ρ=0.815\rho=0.815 at 99%. This supports the theoretical connection between residual-law estimation and score pivotality (Figure 7).

  • Set efficiency. Lower energy scores also correlate with smaller prediction-set footprints, with ρ=0.794\rho=0.794 at 97% and ρ=0.927\rho=0.927 at 99%. Empirical coverage remains comparable across checkpoints, ranging from 96.75--97.04% and 99.16--99.25%, respectively. These results indicate that better residual-law estimation is associated with tighter prediction sets at comparable coverage (Figure 7).

AuthorsOpenReview ↗

Response to Reviewer FmkN - Part 3

5. Does energy-distance error control pivotality for ellipsoids? (Question 2)

  • New theoretical result. Theorem 14 establishes, under bounded support, uniformly positive definite covariance, Lipschitz Mahalanobis-score CDFs, and uniformly bounded energy-distance retrieval error, that

    Δpiv≤Δell+C(ε+λ).\Delta_{\mathrm{piv}}\leq \Delta_{\mathrm{ell}}+C(\sqrt{\varepsilon}+\lambda).

    Here, ε\varepsilon bounds energy-distance retrieval error, λ\lambda is covariance regularization, and Δell\Delta_{\mathrm{ell}} measures non-pivotality of oracle Mahalanobis scores.

  • Geometry misspecification. The Δell\Delta_{\mathrm{ell}} term captures non-pivotality remaining even with perfect residual-law retrieval, including equal-covariance distributions with different tails. The second term quantifies estimation and regularization error. Since excess expected energy-score risk equals half the energy distance, this result connects our training objective to the downstream pivotality defect for ellipsoids. It does not require the true residual laws to be elliptical (Assumption 13, Theorem 14).

6. Can other scalarizations use the retrieved law? (Question 3)

  • Yes. The constructor Γ\Gamma is independent of representation learning. We evaluate FacCP with OT-CP+, GRCP, and CopulaCP. Learned retrieval improves GRCP efficiency at unchanged coverage, while CopulaCP calibration further reduces FacCP set volume at comparable coverage. OT-CP+ retains its finite-reference limitations (Appendix E.1, Tables 20, 22, 23).

7. Notation clarification (Question 4)

We simplified Sections 3 and 4, using Z=hθ(X)Z=h_\theta(X) for learned retrieval coordinates and S^\widehat S for the induced scalar score, while reserving UU for latent uncertainty-generating variables in controlled experiments. The representation ZZ enters residual-law estimation only through the retrieval weights wj(x)w_j(x).

8. Additional Results

We added real-world evaluations on world models, New York taxi data, and river flow data from Dheur et al. (2025) (Appendix E.2, Figures 9--10, Tables 25--26).

Reviewer_x2aLOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Factorized Conformal Prediction (FacCP) for constructing adaptive multivariate prediction sets around a fixed predictor. FacCP learns a representation for retrieving historical residuals with similar conditional distributions, uses the retrieved residuals to construct query-dependent set geometries, and applies a single global split-conformal threshold. The theoretical analysis connects geometry approximation to score pivotality and approximate conditional coverage under additional assumptions. Experiments on synthetic data, quadcopter rollouts, and Argoverse 2 motion forecasting evaluate coverage and set size, showing improved set efficiency.

Critical strengths

The paper includes extensive synthetic experiments that systematically examine the effects of residual geometry, irrelevant input dimensions, and data and model capacity. These controlled studies help clarify the method’s behavior. The provided code for reproducing the synthetic experiments also supports reproducibility.

Critical weaknesses
  1. Broader real-world validation would be helpful. The experiments provide useful controlled analyses, although validation on real-world data is limited to motion forecasting on Argoverse 2. Given the diversity of uncertainty sources in physical systems, evaluation on additional real-world datasets or tasks would strengthen confidence in the method’s general applicability.

  2. The practical benefit over global CP could be clarified further. On Argoverse 2, FacCP reduces footprint by 14.9% relative to Global SNSA, while whole-path coverage decreases from 95.19% to 94.46%. Comparisons at matched empirical coverage would help clarify how much of the footprint reduction comes from improved geometric adaptation. Reporting computational and memory costs would also help readers assess the benefit relative to the additional training and retrieval requirements.

Questions for authors

Could the authors provide coverage–footprint curves over a range of thresholds for FacCP and Global SNSA on Argoverse 2? This would show the tradeoff between prediction-region size and whole-path coverage across operating points and allow comparison at matched coverage levels.

How much additional training time, inference latency, and memory does FacCP require relative to Global SNSA, including the encoder and residual reference bank?

Could the authors evaluate the method on additional real-world datasets or tasks to assess whether the improvements extend beyond Argoverse 2?

4 comments
AuthorsOpenReview ↗

Response to Reviewer x2aL

We thank the reviewer for the constructive feedback. We have revised the paper with three additional real-world evaluations, coverage--footprint curves at matched empirical coverage for Argoverse 2, and measurements of training, inference, and memory costs. For this reviewer, we provide Highlights\_x2aL.pdf with relevant revisions highlighted for a targeted re-read.

1. Broader real-world validation (Weakness 1, Question 3)

We added three evaluations beyond Argoverse 2 (Appendix E.2, Tables 24--26).

  • World Models. On a learned visual world model trained using real cube manipulation trajectories, FacCP calibrates 60-step latent predictions. It achieves 96.67% whole-trajectory coverage with mean log-volume −1125.31-1125.31, compared with −1060.02-1060.02 (lower is better) for Global SNSA at 98.89% coverage (Figure 9, Table 24).

  • NYC taxi prediction. FacCP supports non-ellipsoidal, potentially disconnected prediction sets. Gaussian splats achieve 95.33% coverage with 22.8% less area than covariance prediction and 41.8% less than resCONTRA. Both baselines achieve 94.00% coverage (Appendix E.2, Table 25, Figure 10).

  • River-flow forecasting. On four correlated hydrological targets, FacCP achieves 95.89% coverage with 42.9% less volume than resCONTRA and 54.6% less than covariance prediction, at comparable coverage (Table 26).

2. Coverage--footprint tradeoff (Weakness 2, Question 1)

  • Matched empirical coverage. For Argoverse 2, we added coverage--footprint curves across 90--99% targets (Figure 6, Appendix D.7). FacCP produces smaller footprints than Global SNSA at matched empirical coverage. At the 95%, 97%, and 99% targets, footprint decreases by 13.7%, 13.8%, and 29.2%, respectively, while empirical coverage differs by only 0.29, 0.15, and 0.44 percentage points. This confirms that the footprint reductions are not explained solely by lower coverage.

3. Computational and memory cost (Weakness 2, Question 2)

  • Training and memory. On one RTX 4090, MAE pretraining takes about 18 minutes and energy fine-tuning takes about 22 minutes. The learned model and residual reference bank occupy approximately 41 MiB of GPU memory (Appendix D.8).

  • Inference latency. Constructing all six 60-step prediction tubes takes 3.615 ms per scene for FacCP versus 0.034 ms for Global SNSA. The additional 3.581 ms includes encoding, retrieval, covariance estimation, and tube construction (Table 15).

  • Further inference latency comparisons. Across additional planar quadcopter benchmarks, FacCP constructs prediction sets in 1.741.74 ms per query, compared with 269.02269.02 ms for CVQR and 374.83374.83 ms for per-timestep OT-CP+. GRCP and CopulaCP are faster but produce substantially larger sets. Together with the 3.6153.615 ms inference time on Argoverse 2, these results demonstrate that FacCP can construct adaptive prediction sets with millisecond-scale overhead, making it suitable for consideration in latency-sensitive applications such as autonomous driving (Appendix D.8, Appendix E.1, Tables 15 and 20--23).

Dear Reviewer x2aL,

Thank you again for your constructive review and thoughtful feedback. We would appreciate your feedback on our three new real-world evaluations, coverage–footprint comparisons at matched empirical coverage on Argoverse 2, and additional measurements of training time, inference latency, and memory overhead.

As the discussion period approaches its end, please let us know if any points require further clarification. We would be grateful if you could consider our responses and additional results in your reassessment.

Kind regards,
Authors

Thanks for the detailed explanation, it addressed my concerns, and I will raise my recommendation.

Dear Reviewer x2aL,

Thank you for your follow-up and for reconsidering your recommendation. We appreciate your consideration of our additional experiments and analyses, and are glad that they addressed your concerns.

Kind regards,
Authors

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

The paper focuses on how to adapt multivariate conformal prediction sets to query-dependent uncertainty, so that their shape changes with the input, while keeping the finite-sample coverage guarantee and calibrating with the full calibration set.

Critical strengths
  1. No retraining: The method does not need retraining of retrieval network to support different set families as a single learned representation can support them, for example ellipsoidal, KDE sets which is practical in robotics.
  2. This work validates it’s proposed FacCP methodology well with real tasks in autonomous driving and quadcopter reachability. It also presents the related work section well with drawbacks of parallel conformal prediction literature and is well written.
Critical weaknesses

Model generalization: How will the model perform in a new or unseen data? Specially if we don’t know what geometries should H_theta residual network supports?

Questions for authors

Address the critical weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer wWZh

We thank the reviewer for the constructive feedback. We have revised the paper to clarify how FacCP generalizes to unseen queries and supports different prediction-set geometries. For this reviewer, we provide Highlights\_wWZh.pdf with relevant revisions highlighted for a targeted re-read.

Our central theoretical insight is that when local geometries approximate oracle regions containing the same conditional probability mass, the resulting nonconformity scores become approximately comparable across queries (Theorem 3). To realize this principle, FacCP learns a latent space where covariates with similar conditional residual distributions are close, using a strictly proper energy-score objective. The retrieved residual law then determines the local geometry, independently of the chosen set family.

1. Generalization to unseen queries (Weakness 1, Question 1)

  • New queries and validity. All reported evaluations use held-out queries. For an unseen query xx, the learned representation hθ(x)h_\theta(x) determines retrieval weights over reference residuals, from which a query-specific geometry is constructed. Theorem 10 guarantees finite-sample marginal coverage under exchangeability, regardless of retrieval accuracy. Theorems 3 and 8 further connect the accuracy of local geometry to local coverage reliability.

  • Distribution shift and new tasks. Corollary 9 bounds coverage loss under covariate shift preserving the conditional response law, without target recalibration. In our controlled scale-shift experiment, FacCP maintains 94.90% coverage versus 87.25% for global geometry at a 95% target. We further demonstrate generalization across tasks through new real-world experiments on learned world models, NYC taxi drop-off prediction, and river-flow forecasting (Appendix E.2, Tables 24--26).

2. Flexibility of the learned representation (Weakness 1, Question 1)

  • One representation, arbitrary set constructors. FacCP learns to retrieve residual distributions rather than parameters of a specific geometry. Any valid nested-set constructor Γ\Gamma can therefore operate on the retrieved law without retraining hθh_\theta (Section 4.3). We demonstrate this with ellipsoids, hyperboxes, KDE sets, and zonotopes, reducing mean log-volume by 2.34--5.57 relative to their global counterparts using the same learned representation (Figure 3A--D).

  • Adapting geometry to the task. Different constructors capture different uncertainty structures. Ellipsoids model correlated residuals, zonotopes provide flexible convex sets, and KDE-based constructions capture non-convex and multimodal regions. On NYC taxi drop-off prediction, Gaussian-splat geometry produces disconnected regions around likely destinations, reducing area by 22.8% relative to covariance prediction (Figure 10, Table 25). While constructor choice affects efficiency and local reliability, finite-sample marginal coverage remains guaranteed under exchangeability (Theorem 10).

Dear Reviewer wWZh,

Thank you again for your constructive review and thoughtful feedback. We would appreciate your feedback on our clarifications regarding generalization to unseen queries, the flexibility of the learned representation across set geometries, and our additional real-world evaluations on learned world models, NYC taxi prediction, and river-flow forecasting.

As the discussion period approaches its end, please let us know if any points require further clarification. We would be grateful if you could consider our responses and additional results in your reassessment.

Kind regards,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.