MT-CC: Multi-Group Temperature Scaling for Asymmetric Calibration Behavior in Class-Incremental Learning

Sungjun Yun; Yonghee Choi; Dong-Jun Han

cs.LGOct 19, 2026

MT-CC: Multi-Group Temperature Scaling for Asymmetric Calibration Behavior in Class-Incremental Learning

Authors: Sungjun Yun, Yonghee Choi, Dong-Jun Han

Organizations: Yonsei University

Abstract

As deep neural networks are required to continuously adapt to evolving data, class-incremental learning (CIL) has become an active research area. However, most existing works primarily focus on improving accuracy while overlooking confidence calibration, which measures the trustworthiness of the model's predictions. Although temperature scaling (TS) is effective for static calibration, it has rarely been explored in continual settings where data distributions evolve over time. In this paper, we introduce Multi-Group Temperature Scaling for Continual Calibration (MT-CC), a novel post-hoc calibration framework that addresses the asymmetric calibration behavior in CIL, where the rate of confidence changes does not align with the rate of accuracy degradation across tasks, leading to different levels of miscalibration. Our key contribution is to design multiple temperature parameters through a lightweight grouping network that separates samples by confidence statistics and task-aware signals in CIL. Furthermore, a Groupwise Difference-between-Confidence-and-Accuracy (GDCA) regularization is incorporated to promote intra-group consistency. Experiments demonstrate that MT-CC significantly improves calibration performance while maintaining accuracy, offering a principled framework for reliable continual learning.

Figures & tables

Figure 1: The reliability diagram visualizes calibration by comparing empirical accuracy and predicted confidence within each confidence bin, while a lower ECE indicates better calibration. (a) In class-incremental learning, accuracy and confidence evolve at different rates across tasks, producing asymmetric calibration behaviors : the gap between confidence and accuracy varies across tasks due to catastrophic forgetting. (b) As a result, a single global temperature cannot simultaneously calibrate these heterogeneous groups. (c) Our MT-CC routes each sample to a group-specific temperature to capture asymmetric calibration behaviors in class-incremental learning.

Figure 2: Confidence-accuracy gap of selected tasks over incremental stages of CIFAR-100. The results reveal asymmetric calibration behaviors in continual learning, where the discrepancy between confidence and accuracy varies across tasks due to catastrophic forgetting. Results for all tasks are provided in Appendix D .

Panel of Figure 3 — Figure 3: Calibration behavior in CIL. (a) The final-stage CIL model is compared with a static model jointly trained on all classes using five equal-frequency logit-margin bins from low to high margin. The confidence–accuracy gap increases with margin only under CIL. (b) Task-wise accuracy and confidence are measured after learning all 10 tasks, showing larger gaps for earlier tasks than for newly introduced tasks.

Panel of Figure 4 — Figure 4: Portion of samples from each class allocated to each group. Results are shown after training on the 10th task of CIFAR-100. Compared to (a) naive confidence-based grouping, which yields imbalanced groups, (b) MT-CC achieves a more effective group allocation, leading to lower ECE.

Replay buffer		5K			10K			20K
CL	CA	ECE( $\downarrow$ )	AECE( $\downarrow$ )	SCE( $\downarrow$ )	ECE( $\downarrow$ )	AECE( $\downarrow$ )	SCE( $\downarrow$ )	ECE( $\downarrow$ )	AECE( $\downarrow$ )	SCE( $\downarrow$ )
ER	Vanilla	$29.95\raisebox{0.0pt}{\scriptsize$ \pm 0.30 $}$	$29.95\raisebox{0.0pt}{\scriptsize$ \pm 0.30 $}$	$2.18\raisebox{0.0pt}{\scriptsize$ \pm 0.02 $}$	$22.46\raisebox{0.0pt}{\scriptsize$ \pm 0.19 $}$	$22.46\raisebox{0.0pt}{\scriptsize$ \pm 0.19 $}$	$1.80\raisebox{0.0pt}{\scriptsize$ \pm 0.02 $}$	$16.39\raisebox{0.0pt}{\scriptsize$ \pm 0.23 $}$	$16.38\raisebox{0.0pt}{\scriptsize$ \pm 0.23 $}$	$1.45\raisebox{0.0pt}{\scriptsize$ \pm 0.02 $}$
	TS [ 12 ]	$29.66\raisebox{0.0pt}{\scriptsize$ \pm 0.33 $}$	$29.63\raisebox{0.0pt}{\scriptsize$ \pm 0.33 $}$	$2.16\raisebox{0.0pt}{\scriptsize$ \pm 0.02 $}$	$22.67\raisebox{0.0pt}{\scriptsize$ \pm 0.17 $}$	$22.65\raisebox{0.0pt}{\scriptsize$ \pm 0.18 $}$	$1.79\raisebox{0.0pt}{\scriptsize$ \pm 0.02 $}$	$16.83\raisebox{0.0pt}{\scriptsize$ \pm 0.18 $}$	$16.80\raisebox{0.0pt}{\scriptsize$ \pm 0.20 $}$	$1.44\raisebox{0.0pt}{\scriptsize$ \pm 0.01 $}$
	ETS [ 50 ]	$30.68\raisebox{0.0pt}{\scriptsize$ \pm 0.25 $}$	$30.65\raisebox{0.0pt}{\scriptsize$ \pm 0.26 $}$	$1.09\raisebox{0.0pt}{\scriptsize$ \pm 0.01 $}$	$23.53\raisebox{0.0pt}{\scriptsize$ \pm 0.16 $}$	$23.53\raisebox{0.0pt}{\scriptsize$ \pm 0.16 $}$	$1.09\raisebox{0.0pt}{\scriptsize$ \pm 0.01 $}$	$17.62\raisebox{0.0pt}{\scriptsize$ \pm 0.22 $}$	$17.58\raisebox{0.0pt}{\scriptsize$ \pm 0.23 $}$	$0.91\raisebox{0.0pt}{\scriptsize$ \pm 0.00 $}$
	PTS [ 42 ]	$30.32\raisebox{0.0pt}{\scriptsize$ \pm 0.44 $}$	$30.32\raisebox{0.0pt}{\scriptsize$ \pm 0.44 $}$	$1.09\raisebox{0.0pt}{\scriptsize$ \pm 0.01 $}$	$22.79\raisebox{0.0pt}{\scriptsize$ \pm 0.36 $}$	$22.80\raisebox{0.0pt}{\scriptsize$ \pm 0.36 $}$	$1.89\raisebox{0.0pt}{\scriptsize$ \pm 0.01 $}$	$16.52\raisebox{0.0pt}{\scriptsize$ \pm 0.27 $}$	$16.51\raisebox{0.0pt}{\scriptsize$ \pm 0.28 $}$	$0.91\raisebox{0.0pt}{\scriptsize$ \pm 0.01 $}$
	T-CIL [ 16 ]	$13.45\raisebox{0.0pt}{\scriptsize$ \pm 0.15 $}$	$13.43\raisebox{0.0pt}{\scriptsize$ \pm 0.13 $}$	$1.79\raisebox{0.0pt}{\scriptsize$ \pm 0.01 $}$	$7.74\raisebox{0.0pt}{\scriptsize$ \pm 0.23 $}$	$7.72\raisebox{0.0pt}{\scriptsize$ \pm 0.24 $}$	$1.49\raisebox{0.0pt}{\scriptsize$ \pm 0.01 $}$	$10.57\raisebox{0.0pt}{\scriptsize$ \pm 0.22 $}$	$10.52\raisebox{0.0pt}{\scriptsize$ \pm 0.24 $}$	$1.27\raisebox{0.0pt}{\scriptsize$ \pm 0.00 $}$

Table 2: ImageNet results with varying replay buffer sizes. Since all evaluated methods are post-hoc, they share the same accuracies of the base models: for ER ,

32.86

% (5K),

41.23

% (10K),

49.52

% (20K); for iCaRL ,

44.08

% (5K),

49.29

% (10K),

54.24

% (20K). Reflecting that SCE values vary with the number of labels, per mille (‰) is used.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

$M$	ECE $\downarrow$	AECE $\downarrow$	Brier $\downarrow$
1 (T-CIL)	$6.03{\pm}0.42$	$6.04{\pm}0.35$	$0.570{\pm}0.005$
2	$\mathbf{4.49{\pm}0.53}$	$\mathbf{4.44{\pm}0.49}$	$\mathbf{0.556{\pm}0.006}$
3	$4.51{\pm}0.47$	$4.48{\pm}0.45$	$0.558{\pm}0.007$
4	$4.50{\pm}0.41$	$4.50{\pm}0.42$	$0.559{\pm}0.007$
5	$4.79{\pm}0.40$	$4.77{\pm}0.41$	$0.561{\pm}0.007$

Appendix

Table 8: Ablation on the number of groups

M

on CIFAR-100 over five seeds.

M=1

corresponds to single-temperature scaling (T-CIL).

Panel of Figure 7 — Figure 7: Comparison of calibration metrics across sequential tasks using CIFAR-100.

Figure 8: Per-task accuracy and confidence on the final step of the CIFAR-100 60-5 split, illustrating the discrepancy between confidence and accuracy across tasks.

	CIFAR-10			CIFAR-100			Tiny-ImageNet
Method	ECE@15 $\downarrow$	ECE@20 $\downarrow$	Brier $\downarrow$	ECE@15 $\downarrow$	ECE@20 $\downarrow$	Brier $\downarrow$	ECE@15 $\downarrow$	ECE@20 $\downarrow$	Brier $\downarrow$
T-CIL	$15.36\raisebox{0.0pt}{\scriptsize$ \pm 1.39 $}$	$15.37\raisebox{0.0pt}{\scriptsize$ \pm 1.40 $}$	$0.498\raisebox{0.0pt}{\scriptsize$ \pm 0.019 $}$	$6.19\raisebox{0.0pt}{\scriptsize$ \pm 0.39 $}$	$6.27\raisebox{0.0pt}{\scriptsize$ \pm 0.38 $}$	$0.570\raisebox{0.0pt}{\scriptsize$ \pm 0.005 $}$	$19.04\raisebox{0.0pt}{\scriptsize$ \pm 0.42 $}$	$19.13\raisebox{0.0pt}{\scriptsize$ \pm 0.41 $}$	$0.883\raisebox{0.0pt}{\scriptsize$ \pm 0.004 $}$
T-CIL+ETS	$14.01\raisebox{0.0pt}{\scriptsize$ \pm 1.31 $}$	$14.05\raisebox{0.0pt}{\scriptsize$ \pm 1.30 $}$	$0.497\raisebox{0.0pt}{\scriptsize$ \pm 0.019 $}$	$5.80\raisebox{0.0pt}{\scriptsize$ \pm 0.45 $}$	$5.87\raisebox{0.0pt}{\scriptsize$ \pm 0.44 $}$	$0.571\raisebox{0.0pt}{\scriptsize$ \pm 0.004 $}$	$22.21\raisebox{0.0pt}{\scriptsize$ \pm 0.21 $}$	$22.25\raisebox{0.0pt}{\scriptsize$ \pm 0.21 $}$	$0.901\raisebox{0.0pt}{\scriptsize$ \pm 0.003 $}$
T-CIL+IRM	$14.90\raisebox{0.0pt}{\scriptsize$ \pm 0.95 $}$	$14.83\raisebox{0.0pt}{\scriptsize$ \pm 1.11 $}$	$0.506\raisebox{0.0pt}{\scriptsize$ \pm 0.019 $}$	$6.88\raisebox{0.0pt}{\scriptsize$ \pm 0.66 $}$	$6.97\raisebox{0.0pt}{\scriptsize$ \pm 0.67 $}$	$0.571\raisebox{0.0pt}{\scriptsize$ \pm 0.006 $}$	$18.08\raisebox{0.0pt}{\scriptsize$ \pm 0.42 $}$	$18.05\raisebox{0.0pt}{\scriptsize$ \pm 0.51 $}$	$0.884\raisebox{0.0pt}{\scriptsize$ \pm 0.005 $}$
T-CIL+AdaTS	$14.31\raisebox{0.0pt}{\scriptsize$ \pm 1.44 $}$	$14.34\raisebox{0.0pt}{\scriptsize$ \pm 1.44 $}$	$0.501\raisebox{0.0pt}{\scriptsize$ \pm 0.023 $}$	$7.76\raisebox{0.0pt}{\scriptsize$ \pm 0.50 $}$	$7.86\raisebox{0.0pt}{\scriptsize$ \pm 0.51 $}$	$0.586\raisebox{0.0pt}{\scriptsize$ \pm 0.004 $}$	$21.75\raisebox{0.0pt}{\scriptsize$ \pm 1.25 $}$	$21.80\raisebox{0.0pt}{\scriptsize$ \pm 1.28 $}$	$0.905\raisebox{0.0pt}{\scriptsize$ \pm 0.012 $}$
MT-CC	$10.22\raisebox{0.0pt}{\scriptsize$ \pm 1.51 $}$	$10.27\raisebox{0.0pt}{\scriptsize$ \pm 1.48 $}$	$0.447\raisebox{0.0pt}{\scriptsize$ \pm 0.015 $}$	$4.62\raisebox{0.0pt}{\scriptsize$ \pm 0.51 $}$	$4.72\raisebox{0.0pt}{\scriptsize$ \pm 0.54 $}$	$0.556\raisebox{0.0pt}{\scriptsize$ \pm 0.006 $}$	$15.87\raisebox{0.0pt}{\scriptsize$ \pm 0.37 $}$	$15.93\raisebox{0.0pt}{\scriptsize$ \pm 0.36 $}$	$0.839\raisebox{0.0pt}{\scriptsize$ \pm 0.003 $}$

Appendix

Table 10: Calibration performance on CIFAR-10, CIFAR-100, and Tiny-ImageNet under different numbers of ECE bins and the multiclass Brier score. Lower values indicate better calibration. Results are averaged over five seeds.

Setting	Single-T ECE $\downarrow$	MT-CC ECE $\downarrow$
Static joint training	$0.87\pm 0.14$	$0.87\pm 0.14$
CIL (task average)	$6.03\pm 0.42$	$\mathbf{4.49\pm 0.53}$

Appendix

Table 11: Comparison of single-temperature scaling and MT-CC in static and class-incremental learning settings on CIFAR-100 over five seeds.

Method	Per-sample inference time ( $\mu$ s)	Relative
TS	1250.41	1.0000 $\times$
PTS	1250.65	1.0002 $\times$
AdaTS	1252.63	1.0018 $\times$
MT-CC	1250.80	1.0003 $\times$

Appendix

Table 13: The average per-sample inference time across all tasks on ImageNet over 3 seeds. Relative time is reported with TS as the baseline (1

\times

). MT-CC requires negligible additional inference overhead compared to standard TS.

Method	Calibration / Routing	Feature	ECE $\downarrow$	AECE $\downarrow$	Brier $\downarrow$
PTS	Instance-wise TS	Top-10 logits	$9.87\pm 0.77$	$10.14\pm 0.67$	$0.596\pm 0.008$
PTS	Instance-wise TS	CIL-aware features	$6.88\pm 0.91$	$6.87\pm 0.91$	$0.575\pm 0.008$
AdaTS	Instance-wise TS	Penultimate feature	$7.67\pm 0.52$	$7.66\pm 0.49$	$0.586\pm 0.004$
AdaTS	Instance-wise TS	CIL-aware features	$6.55\pm 0.74$	$6.52\pm 0.75$	$0.574\pm 0.004$
GC+TS [ 45 ]	Semantic-aware learned grouping + group-wise TS	Penultimate feature	$7.23\pm 0.25$	$7.24\pm 0.26$	$0.576\pm 0.005$
K-means Group-TS	K-means ( $K=2$ ) + group-wise TS	CIL-aware features	$5.41\pm 0.51$	$5.37\pm 0.51$	$0.561\pm 0.005$

Appendix

Table 14: Comparison with feature-matched and grouped calibration methods on CIFAR-100 over 5 seeds. The CIL-aware setting uses the same feature vector as MT-CC.

Setting	b0	b1	b2	b3	b4	b5	b6	b7	b8	b9
	(low)									(high)
Static	1.98	1.99	2.00	1.95	2.00	1.94	1.96	1.92	2.04	2.10
CIL	1.58	1.68	1.77	1.84	2.00	2.13	2.36	2.59	2.99	3.82

Appendix

Table 15: NLL-optimal temperature

T^{*}

across logit-margin bins on CIFAR-100 averaged over 5 seeds.

Setting	Margin region	Low-conf. $T^{*}$	High-conf. $T^{*}$	$\Delta T^{*}$
Static	Low	1.98	1.99	0.01
	High	1.94	2.03	0.09
CIL	Low	1.60	1.95	0.34
	High	2.26	3.32	1.05

Appendix

Table 16: Optimal temperatures after controlling for the logit margin.

$\|S_{\mathrm{grp}}\|$	$[\delta,p_{\mathrm{old}},p_{\mathrm{new}},p]$		$[\delta,p]$		$[\delta,p_{\mathrm{rand1}},p_{\mathrm{rand2}},p]$		$[\delta,p_{\mathrm{old}},p_{\mathrm{new}}]$
$\|S_{\mathrm{grp}}\|$	ECE	Brier	ECE	Brier	ECE	Brier	ECE	Brier
50	$4.77\raisebox{0.0pt}{\scriptsize$ \pm 0.43 $}$	$0.565\raisebox{0.0pt}{\scriptsize$ \pm 0.009 $}$	$4.92\raisebox{0.0pt}{\scriptsize$ \pm 0.50 $}$	$0.573\raisebox{0.0pt}{\scriptsize$ \pm 0.006 $}$	$5.00\raisebox{0.0pt}{\scriptsize$ \pm 0.43 $}$	$0.573\raisebox{0.0pt}{\scriptsize$ \pm 0.006 $}$	$5.98\raisebox{0.0pt}{\scriptsize$ \pm 0.48 $}$	$0.572\raisebox{0.0pt}{\scriptsize$ \pm 0.004 $}$
100	$4.36\raisebox{0.0pt}{\scriptsize$ \pm 0.46 $}$	$0.560\raisebox{0.0pt}{\scriptsize$ \pm 0.007 $}$	$4.69\raisebox{0.0pt}{\scriptsize$ \pm 0.40 $}$	$0.571\raisebox{0.0pt}{\scriptsize$ \pm 0.005 $}$	$4.81\raisebox{0.0pt}{\scriptsize$ \pm 0.45 $}$	$0.571\raisebox{0.0pt}{\scriptsize$ \pm 0.006 $}$	$5.78\raisebox{0.0pt}{\scriptsize$ \pm 0.30 $}$	$0.571\raisebox{0.0pt}{\scriptsize$ \pm 0.005 $}$
250	$4.49\raisebox{0.0pt}{\scriptsize$ \pm 0.45 $}$	$0.558\raisebox{0.0pt}{\scriptsize$ \pm 0.006 $}$	$4.70\raisebox{0.0pt}{\scriptsize$ \pm 0.27 $}$	$0.572\raisebox{0.0pt}{\scriptsize$ \pm 0.005 $}$	$4.73\raisebox{0.0pt}{\scriptsize$ \pm 0.34 $}$	$0.572\raisebox{0.0pt}{\scriptsize$ \pm 0.006 $}$	$5.74\raisebox{0.0pt}{\scriptsize$ \pm 0.33 $}$	$0.570\raisebox{0.0pt}{\scriptsize$ \pm 0.005 $}$
500	$4.54\raisebox{0.0pt}{\scriptsize$ \pm 0.56 $}$	$0.557\raisebox{0.0pt}{\scriptsize$ \pm 0.006 $}$	$4.72\raisebox{0.0pt}{\scriptsize$ \pm 0.41 $}$	$0.571\raisebox{0.0pt}{\scriptsize$ \pm 0.006 $}$	$4.71\raisebox{0.0pt}{\scriptsize$ \pm 0.37 $}$	$0.571\raisebox{0.0pt}{\scriptsize$ \pm 0.006 $}$	$5.83\raisebox{0.0pt}{\scriptsize$ \pm 0.30 $}$	$0.571\raisebox{0.0pt}{\scriptsize$ \pm 0.005 $}$
1000	$4.49\raisebox{0.0pt}{\scriptsize$ \pm 0.53 $}$	$0.556\raisebox{0.0pt}{\scriptsize$ \pm 0.006 $}$	$4.63\raisebox{0.0pt}{\scriptsize$ \pm 0.30 $}$	$0.571\raisebox{0.0pt}{\scriptsize$ \pm 0.007 $}$	$4.59\raisebox{0.0pt}{\scriptsize$ \pm 0.40 $}$	$0.571\raisebox{0.0pt}{\scriptsize$ \pm 0.007 $}$	$5.82\raisebox{0.0pt}{\scriptsize$ \pm 0.31 $}$	$0.571\raisebox{0.0pt}{\scriptsize$ \pm 0.005 $}$

Appendix

Table 17: Benefit of explicitly providing the old/new class structure on CIFAR-100 across different grouping-set sizes. The optimization budget is matched across data sizes and

S_{\mathrm{cal}}

is fixed. Results report ECE (%) and Brier score over five seeds.

Grouping-network strategy	ECE $\downarrow$	AECE $\downarrow$	SCE ( $\times 10^{-3}$ ) $\downarrow$	Brier $\downarrow$
MT-CC	$\mathbf{4.49\pm 0.53}$	$\mathbf{4.44\pm 0.49}$	$\mathbf{5.12\pm 0.11}$	$\mathbf{0.556\pm 0.006}$
Lag-1	$4.82\pm 0.36$	$4.82\pm 0.36$	$5.30\pm 0.12$	$0.558\pm 0.006$
Fresh each task	$4.98\pm 0.35$	$4.96\pm 0.34$	$5.38\pm 0.14$	$0.561\pm 0.007$

Appendix

Table 18: Average calibration performance of different grouping-network update strategies on CIFAR-100 over five seeds.

Condition	G1 retention (%) $\uparrow$	G2 retention (%) $\uparrow$
Classifier and router change	$71.8\pm 7.5$	$42.0\pm 12.4$
Router change only	$94.3\pm 6.3$	$90.3\pm 9.6$

Appendix

Table 19: Group retention across adjacent tasks on CIFAR-100, averaged over five seeds. G1 and G2 denote the lower- and higher-temperature groups, respectively.

Figure 9: Profile-likelihood analysis of perturbed routing solutions at the final task on CIFAR-100 over five seeds. The group temperatures are re-optimized from scratch for each routing perturbation. For each seed, profiled

\Delta\mathrm{NLL}

is measured relative to the minimum profiled NLL across all perturbations, while relative

\Delta\mathrm{ECE}

is measured with respect to the learned routing. The shaded region indicates routing solutions whose profiled NLL lies within

0.1\%

of the seed-wise minimum.

Routing behavior	w/o GDCA	with GDCA
Frequency of minority-group fraction below $20\%$	$20.0\%$	$4.4\%$
Mean minority-group fraction	$29.7\%$	$37.5\%$

Appendix

Table 20: Routing-group sizes with and without GDCA over five seeds. The minority-group fraction denotes the proportion of test samples assigned to the less populated of the two groups.

Dataset	After First Task		After Middle Task		After Final Task
	Acc (%)	Conf (%)	Acc (%)	Conf (%)	Acc (%)	Conf (%)
CIFAR-10	$98.79$	$96.06$	$72.69$	$92.09$	$20.37$	$91.38$
CIFAR-100	$85.62$	$92.60$	$42.38$	$79.14$	$30.84$	$75.08$
Tiny-ImageNet	$70.60$	$88.47$	$9.10$	$68.69$	$4.70$	$69.82$

Appendix

Table 21: Average first-task performance degradation over 5 runs under the CIL setup.

Panel of Figure 10 — Figure 10: Catastrophic forgetting and asymmetric calibration behaviors on CIFAR-100. Each curve tracks a task from its introduction through subsequent incremental stages.

Panel of Figure 11 — Figure 11: (a), (b), (c): Reliability diagrams of T-CIL [ 16 ] using a single temperature $T$ . (d), (e), (f): Reliability diagrams of the proposed MT-CC with optimized grouping. Results are shown after training on the 10th task of CIFAR-100, comparing group 1, group 2, and all classes as determined by the proposed grouping function.

Panel of Figure 12 — Figure 12: Reliability diagrams for CIFAR-100 comparing T-CIL and MT-CC across Tasks 2–10. Overall calibration performance shows consistently strong reliability across most tasks.

Model Calibration

Model calibration aligns predicted confidence or model outputs with observed outcomes and measurements to support more reliable decisions.

See all 214 papers

Quantile Adaptive Temperature Scaling for Confidence Calibration
Jun 19, 2026Omprakash Chakraborty, Leo Fillioux, Ismail Ben Ayed +1Post-Hoc Calibration Probability Calibration
1 ÉTS Montréal, Canada · Université Paris-Saclay, CentraleSupélec, Gustave Roussy, INSERM, CDSU, IHU PRISM, Gif-sur-Yvette
Free-Flow Class-Incremental Learning: Towards Robust CIL under Variable Class Arrivals
Apr 3, 2026Zhiming Xu, Baile Xu, Jian Zhao +2Class-Incremental Learning
National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China · School of Artificial Intelligence, Nanjing University, Nanjing, China · School of Electronic Science and Engineering, Nanjing University, China +1
Temporal Imbalance of Positive and Negative Supervision in Class-Incremental Learning
Mar 2, 2026Jinge Ma, Fengqing ZhuClass-Incremental Learning Catastrophic Forgetting
Purdue University, West Lafayette, Indiana, U.S.A.

Jun 19, 2026J/K move · Enter open · S save

Quantile Adaptive Temperature Scaling for Confidence Calibration

Omprakash Chakraborty, Leo Fillioux, Ismail Ben Ayed, Jose Dolz

1 ÉTS Montréal, Canada · Université Paris-Saclay, CentraleSupélec, Gustave Roussy, INSERM, CDSU, IHU PRISM, Gif-sur-Yvette

Post-Hoc Calibration Probability Calibration Model Calibration Uncertainty Calibration

Open paper AskPDF arXiv

Deep neural networks often produce poorly calibrated confidence estimates, overstating their certainty even when predictions are incorrect. Temperature Scaling remains the most widely used posthoc calibration method due to its simplicity and effectiveness, yet its global, uniform rescaling of logits fails to correct the highly heterogeneous structure of miscalibration observed across the confidence spectrum. In particular, the largest correctness confidence discrepancies arise in different quantile regions depending on the setting, low confidence predictions, where uncertainty matters most, tend to exhibit the largest correctness confidence discrepancies, which standard TS leaves largely unaddressed. We introduce Quantile Adaptive Temperature Scaling (QaTS), a simple and efficient post hoc calibration method that adapts the temperature as a function of a predictions empirical confidence quantile. By mapping confidences into the quantile space, QaTS normalizes the calibration problem, makes the structure of miscalibration explicit and enables a monotone temperature function that adapts across quantiles while leaving well calibrated high confidence predictions largely unchanged. preserving high confidence behavior. This quantile aware formulation aligns naturally with a reparameterized Expected Calibration Error (ECE) objective and yields a sample wise temperature that is robust across a variety of challenging scenarios, such as class imbalance and distributional shifts. Across a broad range of datasets, architectures, evaluation scenarios and diverse tasks, QaTS consistently, and substantially, outperforms state of the art post hoc calibration methods, delivering more reliable and trustworthy confidence estimates without modifying model predictions.

Abstract

Model Calibration

Explore similar work

Quantile Adaptive Temperature Scaling for Confidence Calibration

Free-Flow Class-Incremental Learning: Towards Robust CIL under Variable Class Arrivals

Temporal Imbalance of Positive and Negative Supervision in Class-Incremental Learning