acceptodds
Under review as a conference paper at ICLR 2027

What "Form" of Uncertainty Does Conformal Prediction Quantify?

Abstract

Conformal prediction (CP) is a post-hoc calibration method that returns prediction sets with distribution-free coverage guarantees, and the size of the output set is often interpreted as a notion of uncertainty. We show that the difference in set size between full and split CP is directly connected to Bayesian posterior variance, which is the epistemic uncertainty of the model. Essentially, this difference in set size is an aggregation of two opposite effects: (1) better model accuracy and therefore smaller set size achieved by training over extra calibration points, and (2) the "formability" of the model at test point towards every possible label. The latter effect vanishes with stronger models or more calibration data (as epistemic uncertainty is known for that behavior). It turns out that the model’s formability itself (without the quantization effect of prediction sets, and through estimating the posterior variance) is a useful tool to quantify the epistemic uncertainty. In our empirical evaluation it performs on par with the state of the art in many setups.

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

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HZ71OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates why full conformal prediction can produce larger prediction sets than split conformal prediction. It decomposes prediction margins into calibration-data adaptation, candidate-label “formability,” and calibration-threshold movement. Under exact Bayesian updating, it connects formability to the posterior variance of class probabilities and derives covariance-based margin bounds. It then proposes probability-weighted deterministic refits as a practical uncertainty measure.

Evaluations are on controlled synthetic experiments, exact Bayesian regression, and benchmarks on graphs, images, and multiple-choice language tasks. Results are competitive on image and graph error detection, while language results are weaker.

Critical strengths
  1. The interpretable theoretical connection.

Lemma 3.1 establishes an exact relationship between candidate-label adaptation and posterior predictive disagreement. Theorem 3.2 additionally accounts for calibration-threshold movement. Together, these provide a useful explanation of how refitting affects full CP beyond the usual argument about improved data efficiency.

  1. Complementary theoretical and empirical validation.

The conjugate Bayesian experiment checks the identities and margin bounds directly, while comparisons against pNML, Laplace approximations, ensembles, and other uncertainty measures assess practical usefulness. Competitive image and graph results suggest that the continuous formability score contains useful information beyond the discretized set-size gap.

Critical weaknesses
  1. The main identity does not directly justify the implemented estimator. The theory concerns posterior expectations under exact Bayesian updating, whereas the practical method substitutes deterministic/MAP refits. Without a quantitative connection or direct validation against reference posterior variance, the central epistemic-uncertainty interpretation remains insufficiently supported for the implemented method.

  2. Some CP-efficiency interpretations exceed the established results. The paper describes the calibration-trained reference set as an “adaptation floor” or lower bound. However, the decomposition permits negative net formability when threshold movement exceeds the candidate’s probability increase, and Theorem 3.2 does not generally establish the claimed ordering. Because this reference also lacks coverage validity, its smaller size alone does not establish improved conformal efficiency.

Questions for authors
  1. What quantitative evidence justify interpreting the deterministic score as an approximation to Bayesian posterior variance? Can you provide a direct comparison on a tractable model where both quantities can be computed, rather than relying on error-detection performance?

  2. Under what conditions is the adaptation reference a lower bound on full-CP set size? If it is intended only as an empirical reference, which directional or convergence claims about the full–split gap remain theoretically established?

Reviewer_w79ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the uncertainty quantified by split CP and full CP by comparing the sizes of the prediction sets they generate. The authors introduce the “formability” notion, which measures the difference in the estimated probabilities of the split-CP and full-CP models. The theory also connects this formability to the Bayesian posterior variance. The experiments show that neither split CP nor full CP consistently produces smaller prediction sets, and the gap between their sizes decreases as the training or calibration set increases, or as the dimension of the feature decreases. The experiments also show that using formability as a measure of uncertainty leads to better performance in some setups.

Critical strengths
  1. The theoretical contribution is insightful, specifically, the connection between formability and the Bayesian posterior variance.
  2. The experiments provide strong intuition about the difference between full CP and split CP, and that the gap in their set size depends on the calibration and training data sizes.
Critical weaknesses
  1. The formability used in the experiments is computed through deterministic model refitting or approximated updates. That is, it is not the same as the notion of formability studied in the theoretical analysis. The equation relating formability to posterior predictive variance does not necessarily hold in these settings. Therefore, it is not clear whether the proposed score measures epistemic uncertainty or sensitivity to optimization or measures sensitivity to optimization, regularization, or model refitting.
  2. The practical usefulness of this work is not clear. This work considers two conformity scores for classification tasks, and the analysis does not extend directly to other scores. For example, the experiment in Appendix C.1.6 shows that the APS score behaves differently: split CP produces smaller prediction sets. Therefore, it is not clear if the analysis conducted in this work applies to the broader conformal prediction, or only to specific score functions. Additionally, using formability requires refitting, which might be expensive. Thus, the advantage over cheaper uncertainty estimators is not clear.
Questions for authors
  1. When comparing FCP to SCP, you do not consider conditional coverage. That is, when one method constructs larger sets, it is possible that it achieves better conditional coverage, and this is not measured or addressed in the paper. Can you measure the conditional coverage obtained by each method using conditional coverage metrics?
  2. Can the authors verify empirically how well deterministic formability estimates the Bayesian posterior predictive variance?
  3. How sensitive is the proposed framework to the optimization procedure, initialization, regularization strength, and refitting?
Reviewer_WuzhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces formability, a measure of how much a model’s predicted probability for a candidate label increases after refitting with a pseudo-labeled test point. It uses this measure to explain why full conformal prediction can produce larger prediction sets than split conformal prediction. Experiments further show that a probability-weighted formability score can help detect prediction errors.

Critical strengths

None

Critical weaknesses

I have concerns about the interpretation of the probability-weighted formability score in Eq. (6), line 244. Weighting by baseline probabilities can yield a small score despite substantial relative changes after refitting. Consider two hypothetical binary-classification examples. Let wcw_c denote the baseline probability of class cc and Φc\Phi_c its probability increase after refitting with that candidate label. Compare the paper’s score, U=∑cwcΦcU=\sum_c w_c\Phi_c, with a normalized alternative, R=1c∑cΦc1−wcR=\frac1c\sum_c\frac{\Phi_c}{1-w_c}, where 0<wc<10<w_c<1.

ExampleBaseline probabilitiesCandidate-specific refitted probabilitiesProbability increasesPaper’s score UUNormalized score RR
A(0.99,0.01)(0.99,0.01)(0.99909,0.91)(0.99909,0.91)(0.00909,0.90)(0.00909,0.90)0.0180.0180.9090.909
B(0.50,0.50)(0.50,0.50)(0.70,0.70)(0.70,0.70)(0.20,0.20)(0.20,0.20)0.2000.2000.4000.400

Each candidate-specific refitted probability comes from a separate model, so the two entries need not sum to one. Values for A are rounded.

For A: UA=0.99×0.00909+0.01×0.90≈0.018,U_A=0.99\times0.00909+0.01\times0.90\approx0.018, RA=12(0.009090.01+0.900.99)≈0.909.R_A=\frac12\left(\frac{0.00909}{0.01}+\frac{0.90}{0.99}\right)\approx0.909.

For B: UB=0.50×0.20+0.50×0.20=0.20,U_B=0.50\times0.20+0.50\times0.20=0.20, RB=12(0.200.50+0.200.50)=0.40.R_B=\frac12\left(\frac{0.20}{0.50}+\frac{0.20}{0.50}\right)=0.40.

Refitting realizes approximately 91% of the available probability increase for either candidate label in A, compared with only 40% in B. Nevertheless, the paper’s score ranks B substantially higher even though it is obviously model A is easier to be changed. While this ordering is consistent with the posterior-variance identity, it highlights a distinction between posterior variance and relative susceptibility to candidate labels.

Questions for authors

See weakness.

Comments on the paper

2 comments

We thank the reviewer for pointing out their concerns; in light of the mentioned points we add the following results and discussions, hopefully to address their concerns.

Stability across seeds. On CIFAR-10 (a 10-class image dataset), for 3 data splits and 10 seeds each, we retrained the base model and all 2,000 refits with a new seed and compared the per-image scores between seeds.

ModelRank agreement between two seeds: formabilitySame for entropyError-detection AUROC of formability (spread over seeds)
ResNet-18 features, 64-unit classifier (main model)0.9690.9780.837 (+- 0.014)
ResNet-50 features, 64-unit classifier0.9860.9870.897 (+- 0.010)
ViT-B/16 features, 64-unit classifier0.9870.9870.925 (+- 0.008)

Refits. For one test point and one candidate class, a refit trains a new network from freshly initialised weights on all labelled data plus that one pseudo-labelled point. The base model and every refit of the same data split start from the same initial weights and use the same dropout, so they differ only by the one added label. On images the network is a two-layer classifier with 64 hidden units on top of fixed image features, trained for 300 passes over the data with the Adam optimiser.

HMC. We evaluated over 3 splits of 200 points for CIFAR-10 and CoraML datasets and the AUROC pooled over the 600 points; the mean of the three per-split AUROCs is the same to within 0.003.

ScoreCIFAR-10CoraML
HMC variance0.8310.802
Formability (refit)0.8310.802
Softmax Entropy0.8340.816

Comparison with softmax entropy. We note that on the real-data the amount of epistemic uncertainty is not clear; specially in current baselines where the models are already performing well. The main comparison of softmax entropy and formability is where we can see each of the uncertainties separately. Fig 4. specifically shows that when blobs get closer formability reacts to the decrease of the epistemic (at the center) uncertainty while softmax remains steady.

We added this experiment which further highlights this effect. We synthesize a dataset where the input is a one-hot indicating either of the 3000 groups, and each group is assigned with a n, and \epsilon, which controls the number of observation, and the true posterior distribution which is a permutation of (1−ϵ,ϵ/2,ϵ/2)(1 - \epsilon, \epsilon/2, \epsilon/2). The first controls the epistemic and the second the aleatoric uncertainty. The model is a linear model so that Bayesian posterior can be computed analytically. Rank agreement with the exact posterior variance is 1.00 (1 is perfect) for formability, and 0.48 for softmax entropy. This is the breakdown in different regions of high/low alea/epis uncertainties.

RegionFormabilitySoftmax entropy1 − max softmax prob.
Few labels, clean labels (epistemic high, aleatoric low)0.96*0.930.92
Few labels, noisy labels (both high)0.98*0.270.14
Many labels, clean labels (both low)1.000.760.77
Many labels, noisy labels (epistemic low, aleatoric high)1.000.200.22
All 3,000 contexts1.000.480.49

Computational cost. While we agree the method is costly, we can reduce this cost by only applying the refit to a small part of the network, like the last layers, or only the weights with higher magnitude. The following compare different refit regimes in comparison with a 5 member ensemble (CIFAR-10, 200 images):

MethodTrainingsEpochs eachCost (full trainings)Cost relative to a 5-member ensembleError-detection AUROC
Softmax entropy (base model only)130010.2×0.844
Ensemble variance530051×0.836*
Formability, full refit1 + 2,0003002,001400×0.842
Formability, all weights from the base model1 + 2,000300 + 5033467×0.838
Formability, random 2% of weights1 + 2,000300 + 5033467×0.832
Formability, largest 2% of weights1 + 2,000300 + 5033467×0.824
Formability, smallest 2% of weights1 + 2,000300 + 5033467×0.761
Formability, last layer only, features computed once1 + 2,000300 + 507.51.5×0.823

We discuss the points on full cp in the next comment.

Full CP itself. We don't claim that full cp recovers conditional coverage. In fact it doesn't and on this metric it is as bad as Split CP. What we claim is that formability can detect those wrong sets. The table below is on CoraML/GCN

ScoreRegionCoverage, fullCoverage, splitSet size, fullSet size, split
TPSAll test nodes0.8990.9022.941.21
TPSWorst slab0.8850.8852.801.23
TPSRest0.9010.9032.941.21

The binarization effect. First we should mention that the full-cp set size itself has a higher than chance AUROC; the set size difference works with a below-chance performance (results on CIFAR-10/ResNet):

ScoreAUROC
Set-size gap, full − split0.453 ± 0.033
Full CP set size0.664
Split CP set size0.703
Softmax entropy0.853
Formability0.854

Again note that softmax entropy, and split set size are encoding aleatoric uncertainty as well, so we can not entirely see this as epistemic effect. But to highlight how binarization hurts the performance we replace the hard margin rule with a sigmoid function \simaτ(m)=1/(1+exp(−m/τ))\sima_\tau(m) = 1 / (1 + exp(−m / \tau)) where decreasing τ\tau sharpens the decision towards a binary margin. Here is the result:

Threshold width τ10.30.10.050.01hard
AUROC of the softened gap0.7210.7150.5800.4640.4440.453

As shown, increasing the sharpness of the margin results in an exact reverse effect, while the soft sigmoid margin itself can still show a relatively high AUROC.

LLMs. We agree that the LLM result remains a limitation; indeed we intentionally brought that result to highlight the limitation of the framework. What we believe for the reason is that formability requires the model to be at MAP, and be able to move to another refit MAP within a refit. This means that if the optimizer has not landed in an optimum, the metric can not compete with other methods. The same assumption holds for Laplace approximation, and similarly it also fails for LLMs (worse than formability).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.