acceptodds
Under review as a conference paper at ICLR 2027

Towards Reliable Uncertainty Quantification: When Conformal Prediction Meets Class Incremental Learning

Abstract

Reliable uncertainty quantification is critical for Class-Incremental Learning (CIL), making Conformal Prediction (CP) a promising approach due to its finite-sample coverage guarantees. However, under a fixed calibration budget, CP must allocate limited calibration samples across classes, while discarded samples from old classes cannot be recovered. As class-wise calibration demands evolve over time, allocations tailored to current CP reliability may become inadequate for future tasks. To address these challenges, we propose BACA, a Budget-Aware Calibration Allocation framework that improves CP reliability in CIL scenarios. Specifically, BACA contains three complementary components: (1) Current Demand Estimation estimates class-wise calibration demand by characterizing the sensitivity of prediction-set instability to calibration size; (2) Irreversibility-Aware Future Protection identifies classes vulnerable to irreversible future risk and protects them from excessive calibration-sample removal; and (3) Marginal Budget Allocation, which reallocates the fixed calibration budget by balancing the marginal stability benefit of adding samples and the marginal instability cost of removing samples. Extensive experiments on three benchmarks with four CIL learners show that BACA consistently improves prediction-set stability under fixed calibration budgets. It achieves higher Jaccard similarity and lower disagreement than uniform allocation while maintaining comparable coverage and average prediction-set size.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Conformal Prediction, Class Incremental Learning, Uncertainty Quantification, Budget-Aware Calibration Allocation

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_R9LVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies \emph{calibration-memory allocation} for conformal prediction (CP) in class-incremental learning (CIL). The authors observe that under a fixed memory budget, calibration samples of old classes cannot be recovered once discarded, while the calibration demand is heterogeneous across classes and evolves over tasks. Existing work either assumes sufficient calibration resources or focuses on constructing conformal uncertainty estimates, leaving the problem of calibration-memory allocation under a fixed budget unexplored. The problem is formalized as the constrained optimization in Eq. (4) (a fixed total budget McalM_{\mathrm{cal}} together with the irreversibility constraint Cct⊆Cct−1C_c^{t}\subseteq C_c^{t-1}), and three theoretical results are provided:Theorem 1 (finite-calibration disagreement law Dc(m)=ac/mD_c(m)=a_c/\sqrt{m}), Theorem 2 (oracle allocation mc⋆∝ac2/3m_c^{\star}\propto a_c^{2/3}, with equal allocation optimal only when all aca_c coincide), and Theorem 3 (a lower bound on irreversibility-induced future risk). Building on these, the authors propose \paper{}, a three-stage framework (Current Demand Estimation, Irreversibility-Aware Future Protection, Marginal Budget Allocation), and evaluate it on CIFAR-10/100 and Tiny-ImageNet with four CIL learners (ER, DER, MEMO, TagFex). Relative to Equal Allocation, \paper{} improves prediction-set stability (Jaccard, Disagreement) while leaving coverage and average set size essentially unchanged.

Critical strengths

Well-chosen and cleanly formalized problem. The two constraints in Eq.(4) a fixed total and a nested (monotone) sample store, capture the real tension in fixed-memory CIL, and are structurally the same as those of a replay buffer. The problem statement in Sec.3 is unambiguous.

Tight theory-to-algorithm correspondence. The three theorems map one-to-one onto the three stages (Thm. 1 -> Stage 1, Thm. 3-> Stage 2, Thm. 2 -> Stage 3), so this is not a method first, theory bolted on later paper. Theorem 2 yields a falsifiable claim: equal allocation is oracle-optimal only if all aca_c are equal. The 3/2 transformation is elegant. The authors transform the protected target support back into the demand space using the following formula: a~c=a^c (mctar/mc(0))3/2\tilde a_c=\hat a_c\,(m_c^{\mathrm{tar}}/m_c^{(0)})^{3/2}, making the third stage still a single allocation process.

Critical weaknesses

The authors claim that "different finite calibration samples may produce different prediction sets... These differences can change downstream decisions and make uncertainty quantification less trustworthy." However, the experiments in the entire paper (Table 1, Table 2, Fig. 3, Appendix C) measure only four internal CP metrics: Jaccard, Disagreement, Coverage, and Set Size. There is no experiment that measures the quality of any downstream decision. In other words, the paper claims that differences in prediction sets will change downstream decisions, but never demonstrates the consequences of those differences.

The paper treats McalM_{\mathrm{cal}} as standing alongside an exogenous replay budget, yet never states the relationship between the two. Eq. (4) constrains only ∑cmct=Mcal\sum_c m_c^{t}=M_{\mathrm{cal}}; in the experiments McalM_{\mathrm{cal}} is an independently given constant (Eq. (63) in Appendix C.1) that does not participate in gradient updates. The text never imposes a constraint of the form Mcal+Breplay=constM_{\mathrm{cal}}+B_{\mathrm{replay}}=\text{const}, nor does it clarify whether CctC_c^{t} is disjoint from the replay buffer. Two consequences follow. (i) If McalM_{\mathrm{cal}} is additional memory, this contradicts the paper's own fixed-memory premise ("unconstrained retention would make memory consumption grow continuously"). (ii) If McalM_{\mathrm{cal}} is carved out of a fixed total, then every sample moved to calibration is one removed from replay, and accuracy must necessarily be affected. That is precisely the experiment capable of answering "is it useful for continual learning?", and it is omitted entirely.

Sec. 6.1 and Appendix C.1 state that Equal Allocation is the only baseline, on the grounds that "existing methods do not directly address calibration-budget allocation." That reasoning explains the absence of an off-the-shelf method for the same task, but it does not explain why existing allocation rules were not brought in as ablations. Using ER/DER/MEMO/TagFex partially addresses the training side and, to some extent, the demand for a comparison against replay methods; strictly speaking, however, the most directly opposed baselines are the memory-management strategies already established in the CIL literature (iCaRLherding, GDumb greedy balancing, Class-Balanced allocation) repurposed as rules for how many calibration samples each class receives.

Questions for authors

Q1. What exactly is the relationship between the calibration memory McalM_{\mathrm{cal}} and the replay buffer BreplayB_{\mathrm{replay}}? Are they disjoint, i.e. are calibration samples partitioned out of the replay samples? If total memory is fixed, what is the accuracy cost of assigning memory to calibration?

Q2. Suppose part of McalM_{\mathrm{cal}} is returned to training (for example, halving McalM_{\mathrm{cal}} and doubling replay). Does the stability advantage still hold? Has this trade-off ever been attempted?

Q3. Please provide evidence on at least one downstream metric of the class-incremental learning task itself (for example, average accuracy at any point in time) showing that improved prediction-set stability translates into better decisions.

Q4. If each class is given a fixed per-class cap instead of an equal share, does the method still show a significant advantage? This bears directly on the practical value of Theorem 2.

3 comments
AuthorsOpenReview ↗

Further Clarifications on Practical Implications and Calibration Memory Management

W1 and Q3

We thank the reviewer for raising this concern. Papamichalis et al. (2026) identify three practical consequences of prediction-set inconsistency : (1) hospitals may fail to establish a consistent predictive system for joint deployment; (2) regulators may be unable to verify whether a vendor's deployed system matches its audited version; and (3) inconsistent LLM answer sets may lead to unreliable cache reuse. Their experiments report changes in answer sets across independent calibrations for 45% of Llama-3.2-1B contexts and 76% of GPT-2 contexts. Furthermore, Cresswell et al. (ICML 2024) show that prediction sets affect human decision accuracy, while their ICLR 2025 study demonstrates that equalized coverage does not prevent disparate impacts on downstream decisions. These findings highlight the practical consequences of prediction-set inconsistency beyond coverage and set size.

Regarding ACC, BACA only reallocates calibration samples without modifying model training or parameters. Thus, its benefit is more consistent uncertainty information rather than improved classification accuracy.

W2 and Q1

We thank the reviewer for raising this concern. At initialization, we divide the fixed-size memory buffer into two fixed-size, non-overlapping sub-buffers: the CIL Replay Buffer and the CP Calibration Memory. The former retains representative training samples, while the latter stores randomly sampled calibration data independent of the training and test sets.

This separation is necessary because the two buffers serve fundamentally different purposes. The Replay Buffer supports model training and knowledge retention, whereas the Calibration Memory is exclusively used to estimate conformal quantiles and construct prediction sets, without participating in model training. In standard CP, random sampling is required to maintain exchangeability between calibration and test samples.

This design allows both buffers to serve their respective purposes within the fixed total memory budget. However, jointly optimizing their allocation under a shared budget is an interesting direction for future work, beyond the scope of this study. We also evaluate BACA under different calibration memory sizes, demonstrating its effectiveness across varying calibration budgets.

W3

We thank the reviewer for the suggestion. However, our objective differs fundamentally from CIL replay-buffer management. Replay buffers select informative and representative samples to improve model learning, whereas CP calibration memory is intended to reflect the test-data distribution. Therefore, calibration samples must be randomly selected rather than manually prioritized to preserve the exchangeability required by standard CP. Accordingly, BACA does not select individual samples; instead, it determines the number of calibration samples allocated to each class, with random sampling within each class. Therefore, existing CIL replay methods, which focus on selecting informative training samples, are not directly comparable to BACA.

Q2

Reducing the calibration memory may weaken the stability advantage of BACA. Nevertheless, BACA can retain its advantage over equal allocation under reduced calibration budgets, as Stage 3 adaptively distributes the available samples according to class-wise calibration demand and future protection requirements. When the budget becomes extremely small, BACA may approach equal allocation due to limited allocation flexibility, and its advantage may consequently diminish. Our experiments under different calibration memory sizes further evaluate this advantage across varying budgets.

Q4

We thank the reviewer for this question. We consider two settings of fixed per-class caps.

First, assigning the same fixed cap (X) to every class has two cases: (1) without a total budget constraint, the calibration memory grows as new classes arrive, violating our fixed-budget setting; (2) under a fixed total budget, fully utilizing all class caps reduces to equal allocation, while leaving some caps unused leads to budget underutilization.

Second, assigning different fixed caps (XcX_c) to individual classes requires prior knowledge of their respective calibration demands. However, these demands evolve as new classes arrive, making such predefined caps difficult to determine.

In contrast, BACA fully utilizes the fixed calibration budget through adaptive class-wise allocation, prioritizing classes with greater current calibration demands and future protection needs. This enables BACA to accommodate evolving class-wise requirements without relying on predefined per-class caps or uniform allocation.

Reviewer_R9LVOpenReview ↗

rebuttal improves clarity but does not substantially address the issues

I appreciate the authors’ detailed rebuttal and the clarification that replay memory and calibration memory are disjoint partitions of a fixed total memory budget. However, this clarification ultimately reinforces my main concern: every calibration sample necessarily comes at the expense of a replay sample, yet the paper does not evaluate this trade-off or show whether the stability gains justify the loss of replay capacity.

More importantly, the paper motivates BACA through downstream reliability and decision consistency, but the experiments only report CP-internal metrics such as Jaccard similarity and disagreement. While the rebuttal cites prior work showing why prediction-set instability may matter, it does not demonstrate any downstream benefit in the continual-learning setting considered here.

Overall, the rebuttal improves clarity but does not substantially address the practical significance or experimental completeness. Therefore, I prefer to maintain my score at 2.

AuthorsOpenReview ↗

Clarifying the Memory Trade-off and Practical Significance of Stability Improvements

We thank the reviewer for raising these important concerns.

Regarding the memory trade-off, BACA does not require additional calibration memory or reduce replay capacity compared with Equal Allocation. Both methods operate under the same fixed total memory budget and identical preallocated Replay Buffer and Calibration Memory sizes. They differ only in how the fixed calibration budget is distributed across classes. Therefore, BACA's stability gains are achieved without sacrificing additional replay capacity.

Importantly, our contribution is to improve calibration memory utilization through adaptive class-wise allocation under a fixed budget, rather than to increase memory consumption. BACA considers both current class-wise calibration demand and irreversibility-aware future protection when allocating calibration samples. Compared with Equal Allocation, BACA improves Jaccard similarity by 3.54 percentage points and reduces prediction-set disagreement by 7.76 percentage points on average, while maintaining comparable empirical coverage and prediction-set size. These improvements are achieved through more effective allocation of the existing calibration budget, without additional memory consumption.

Regarding downstream benefits, improved prediction-set stability can make subsequent decisions less sensitive to random calibration sampling. For example, consider an industrial monitoring system that determines inspection priorities based on prediction sets. For the same observation, different calibration subsets may produce different candidate sets, such as {Normal,Minor Fault\text{Normal},\text{Minor Fault}} and {Minor Fault,Critical Fault\text{Minor Fault},\text{Critical Fault}}. Under a policy that triggers urgent inspection whenever a critical fault is among the candidates, these outputs lead to different inspection priorities. By reducing such prediction-set variations, BACA can promote more consistent risk assessment, reduce unnecessary changes in inspection priorities, and make inspection resource allocation more predictable. Such consistency is also beneficial for reproducible risk assessment and operational auditing.

We acknowledge that quantifying the actual utility of these benefits requires application-specific decision policies and cost models, which we leave for future work.

Reviewer_2jJ7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to allocate a fixed, irreversible calibration memory across classes in class-incremental learning. It proposes BACA, which estimates class-wise calibration demand, protects classes judged vulnerable to future demand increases, and reallocates samples using marginal stability gains and losses. experiments on three image benchmarks and four CIL learners report improved prediction-set stability over uniform allocation.

Critical strengths

Relevant problem formulation. The paper identifies a concrete tension between the fixed memory budget, heterogeneous class-wise calibration needs, and the irreversibility of deleting old-class samples.

The evaluation and separate component analyses are broad.

Critical weaknesses

The coverage guarantee is not established for the full continual-learning setting. The paper correctly associates conformal coverage with exchangeability, but its experiments report empirical coverage rather than proving that coverage remains valid after the model is updated and calibration memory is selectively retained or deleted.

The future-protection component depends on a modeling choice and tuned parameters. The future-risk score is based on a bounded uncertainty model for demand growth, which the paper explicitly says is not implied by conformal coverage theory.

The main baseline is uniform allocation, so the results do not show how BACA compares with other plausible demand-aware or predictive allocation strategies.

Questions for authors

See the weakness section.

1 comment
AuthorsOpenReview ↗

Clarifications on Calibration Validity, Future Protection, and Allocation Baselines

W1

We thank the reviewer for raising this concern. At each incremental step, BACA uses the updated model to construct prediction sets and recalculates calibration quantiles using retained calibration memory. Thus, calibration always corresponds to the current model rather than outdated quantiles.

BACA determines class-wise sample quantities rather than selecting individual samples based on their scores or importance. Samples are then uniformly randomly selected within each class. This avoids explicit score-based sample prioritization and preserves random sampling within each class.

Our experiments also evaluate empirical coverage after each update, showing that BACA maintains comparable coverage while improving prediction-set stability.

W2

We thank the reviewer for this concern. Future Protection involves a modeling choice and two control parameters, but its effectiveness does not rely heavily on specific parameter choices. Predicting future demand for each class is challenging because future tasks are unavailable, newly arriving classes have limited historical observations, and maintaining class-specific predictors becomes increasingly costly as classes grow. We therefore use a shared, data-dependent uncertainty scale based on current class-wise demand, with ρ\rho and η\eta controlling protection scope and strength. Sensitivity results across 15 configurations support this design: Jaccard varies from 70.34% to 70.72%, disagreement from 85.42% to 85.91%, and coverage from 87.53% to 87.66%.

The bounded uncertainty model, 0≤Δcτ≤aˉt0\leq\Delta_c^\tau\leq\bar a^t, addresses unknown future demand growth under irreversible memory constraints. Without a bound, a finite worst-case future-risk characterization cannot be obtained. Using current average class-wise demand as a shared scale provides a tractable risk criterion without requiring class-specific future-demand prediction.

Importantly, this modeling assumption is separate from CP's coverage assumptions and does not modify the CP calibration procedure. CP coverage concerns exchangeability between calibration and test samples, whereas our model characterizes future demand growth under irreversible memory constraints. It addresses the complementary problem of allocating limited calibration memory, rather than establishing an additional coverage guarantee.

W3

We thank the reviewer for this suggestion. Equal allocation is a natural baseline because our benchmark datasets are class-balanced and the total calibration budget is fixed. It also avoids additional assumptions about class-wise demand.

Existing CIL memory management methods mainly select representative or informative training samples for replay, rather than determine class-wise sample quantities. These strategies cannot be directly transferred to CP calibration memory, which must preserve exchangeability between calibration and test samples. Thus, they do not provide directly comparable allocation baselines.

We also considered conventional allocation strategies. Proportional allocation based on class sizes is unsuitable for CIL, since old-class calibration samples are restricted by retained memory, whereas new classes provide additional candidates. Sample availability therefore reflects class arrival and memory constraints rather than calibration demand.

We further evaluated Neyman Allocation, which distributes samples according to class-wise nonconformity-score standard deviations. It performs worse than Equal Allocation in both stability metrics: Jaccard decreases by 5.65 and 6.12 pp, while disagreement increases by 2.79 and 3.87 pp under α=0.1\alpha=0.1 and 0.20.2, respectively. BACA exceeds Neyman Allocation by 9.03 and 9.70 pp in Jaccard and reduces disagreement by 9.82 and 12.69 pp, respectively.

This difference arises from their distinct objectives. Neyman Allocation minimizes stratified-estimator variance, whereas our objective is prediction-set stability. Its variance-based allocation does not directly capture prediction-set disagreement, explaining its inferior performance even compared with Equal Allocation. Therefore, we report this comparison in the rebuttal but do not include Neyman Allocation as a formal baseline, since it targets a different optimization objective.

Reviewer_Hkh5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies the problem of class-incremental learning under a memory budget, in which at each timestep, a learner is provided data for a new set of classes it has not seen previously, and is evaluated on its ability to predict over the union of all classes it has seen thus far. Because this process happens over many timesteps, the amount of training data increases with each timestep and may become too much for the learner to store. Previous work has studied a memory-budgeted model of class-incremental learning where the learner can only store in memory a small budget of datapoints seen at previous timesteps, and must discard the rest.

The authors consider the problem of computing conformal prediction sets in this memory-budgeted CIL setting, and observe that prediction set stability becomes an issue as the data available from earlier classes shrinks over time to accommodate newer classes. The authors present a set of theoretical results characterizing the stability of sets, optimal budget allocation, and the potential regret of forgetting points that are needed in the future. They use these results to propose a framework for deciding which points to allocate toward the memory budget at each time step, and demonstrate the performance of their method empirically.

Critical strengths
  • The question of how best to allocate a limited memory budget for calibration data is interesting and seems well-motivated in the CIL setting.
  • The authors present extensive experiments evaluating their method over four different CIL methods and coverage levels, and their method consistently improves over a naive allocation approach when evaluated on two different disagreement metrics.
Critical weaknesses
  • In general, the paper is quite difficult to follow, and would have benefited from a clearer description of the data access model as well as made sure all terms are properly defined in the main body. A number of terms such as q_c are used without introduction and only defined during proofs in the appendix.
  • I would have appreciated more discussion and intuition for the characterizations provided by the theoretical results in Section 4. For instance, on what sorts of class-conditional distributions do we expect higher/lower values of a_c?
  • The experimental results consistently show improved prediction set stability under BACA, but the improvements are quite minimal in many cases. It would be useful to include information about the variation in stability across tasks rather than just on average.
Questions for authors

It is a bit difficult to track what data the choices of BACA depend on at each time step. Can the authors speak to how the budget allocations depend on the saved data, and whether this introduces any sort of selection bias over time?

2 comments
AuthorsOpenReview ↗

Clarifications on Data Access, Theoretical Insights, and Task-wise Stability

W1

We thank the reviewer for this suggestion. We revised Algorithms 1–4 to clarify each stage's inputs, outputs, and data dependencies. At each task, old classes use only previously retained calibration samples, whereas new classes use their current candidate pools. Stage 1 estimates class-wise demand, Stage 2 assesses irreversible future risk, and Stage 3 allocates the fixed budget and updates memory. The revised pseudocode details this data flow.

We also clarify that qcq_c is the population (1−α)(1-\alpha)-quantile of the class-conditional nonconformity-score distribution; fc(qc)f_c(q_c) and hc(qc)h_c(q_c) are the calibration- and deployment-score densities, respectively, evaluated at this quantile.

W2

We thank the reviewer for requesting further intuition.

Theorem 1 (Class-wise demand). The coefficient

ac=2α(1−α)πhc(qc)fc(qc)a_c=2\sqrt{\frac{\alpha(1-\alpha)}{\pi}}\frac{h_c(q_c)}{f_c(q_c)}

characterizes the sensitivity of prediction-set membership to finite calibration samples. A smaller fc(qc)f_c(q_c) means fewer calibration scores lie near the target quantile, making its estimate more variable. A larger hc(qc)h_c(q_c) means more deployment scores lie near the prediction-set boundary, so threshold fluctuations affect more predictions. Thus, sparse calibration density or dense deployment density near qcq_c leads to larger aca_c and greater calibration demand.

Theorem 2 (Fixed-budget allocation). The leading-order total variation is J(m;a)=∑cac/mcJ(\mathbf m;\mathbf a)=\sum_c a_c/\sqrt{m_c}. Minimizing it under a fixed budget gives mc⋆∝ac2/3m_c^\star\propto a_c^{2/3} when allocation bounds are inactive. Hence, classes with greater instability demand should receive more samples; equal allocation is optimal only when their demand coefficients are identical. The M−1/2M^{-1/2} rate also indicates diminishing returns as the total budget grows.

Theorem 3 (Future risk). Under evolving demand and irreversible deletion, future exposure is

Ecτ(m)=Δcτ[1m−1mcref]+.E_c^\tau(m)=\Delta_c^\tau\left[\frac{1}{\sqrt m}-\frac{1}{\sqrt{m_c^{\mathrm{ref}}}}\right]_+.

It increases with future demand growth and with the reduction below the reference support. Deleted samples cannot be recovered. Thus, allocation based only on current demand may leave classes vulnerable to future demand growth, motivating BACA's future protection.

W3

We thank the reviewer for this suggestion. We therefore compared BACA with Equal Allocation task by task, averaging paired differences over learners, budgets, and α∈{0.1,0.2}\alpha\in\{0.1,0.2\}. Positive values indicate improvement by BACA.

DatasetTasksJaccard gain (pp)Disagreement reduction (pp)
CIFAR-101–5+0.049; +0.106; +0.128; +0.153; +0.147-0.049; +0.160; +0.265; +0.398; +0.273
CIFAR-1005–10+6.235; +6.842; +6.410; +6.695; +5.241; +5.252+17.985; +14.353; +17.410; +14.323; +14.809; +9.721
Tiny-ImageNet5–10+5.456; +3.569; +4.335; +3.144; +3.479; +3.025+9.395; +2.941; +8.917; +3.027; +6.601; +1.969

BACA improves both metrics consistently across all tasks on CIFAR-100 and Tiny-ImageNet, while gains on CIFAR-10 are small and disagreement slightly increases in Task 1. Thus, the magnitude of improvement varies by dataset and task, rather than being uniform. Average coverage changes are +0.11, +0.18, and +0.18 pp for CIFAR-10, CIFAR-100, and Tiny-ImageNet, respectively, indicating no systematic coverage reduction.

Q1

At each task, BACA uses only the currently available calibration pools: previously retained samples for old classes and current candidate samples for new classes. Stage 1 estimates demand from repeated random calibration-subset pairs; Stage 2 uses this demand and previous memory to assess irreversible risk; Stage 3 determines class-wise retention quantities under the fixed budget.

BACA adapts the number of samples per class, not the selection priority of individual samples. After allocation, samples are randomly retained within each class, so each sample in a class's available pool has the same selection probability at that update. Thus, BACA introduces class-wise differences in retained quantities but does not rank or preferentially retain individual samples based on their nonconformity scores.

Thanks for your detailed rebuttal! It has addressed some of my concerns, but I still have reservations about the overall clarity and the other reviewers have brought up some reasonable concerns, so I will maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.