When and Where to Reset Matters for Long-Term Test-Time Adaptation

Taejun Lim; Joong-Won Hwang; Kibok Lee

cs.LGMar 4, 2026

When and Where to Reset Matters for Long-Term Test-Time Adaptation

Authors: Taejun Lim, Joong-Won Hwang, Kibok Lee

Organizations: 1Yonsei University · 2ETRI

Abstract

When continual test-time adaptation (TTA) persists over the long term, errors accumulate in the model and further cause it to predict only a few classes for all inputs, a phenomenon known as model collapse. Recent studies have explored reset strategies that completely erase these accumulated errors. However, their periodic resets lead to suboptimal adaptation, as they occur independently of the actual risk of collapse. Moreover, their full resets cause catastrophic loss of knowledge acquired over time, even though such knowledge could be beneficial in the future. To this end, we propose (1) an Adaptive and Selective Reset (ASR) scheme that dynamically determines when and where to reset, (2) an importance-aware regularizer to recover essential knowledge lost due to reset, and (3) an on-the-fly adaptation adjustment scheme to enhance adaptability under challenging domain shifts. Extensive experiments across long-term TTA benchmarks demonstrate the effectiveness of our approach, particularly under challenging conditions. Our code is available at https://github.com/YonseiML/asr.

Figures & tables

Figure 1: Illustrative comparison between a naive reset approach (RDumb; Press et al. (2023) ) and our Adaptive and Selective Reset (ASR) based on the same model (ETA; Niu et al. (2022) ). RDumb fully resets parameters at fixed intervals (e.g., every 1000 steps), whereas ASR dynamically decides when and where to reset, achieving more stable (i.e., less abrupt performance drop at each reset) and higher performance. Dotted vertical lines indicate when resets occur.

Figure 2: Overview of our Adaptive and Selective Reset (ASR) scheme, which compares prediction concentration $\mathcal{C}_{t}$ with its cumulative counterpart $\bar{\mathcal{C}}_{t-1}$ for each test batch from a long domain stream, triggers a reset when $\mathcal{C}_{t}>\bar{\mathcal{C}}_{t-1}$ , indicating that the model is corrupted severely enough to collapse, and determines layers to reset based on $\mathcal{C}_{t}-\bar{\mathcal{C}}_{t-1}$ , which reflects how severely the model is corrupted. On the upper side, icons inside dashed boxes, labeled with numbers, denote class labels. White icons represent correct predictions, while black icons represent incorrect predictions.

Figure 3: Corr. of

\mathcal{C}_{t}

and Acc.

Method	Easy	Medium	Hard	Mean
ETA (ICML’22)	43.24 $\pm$ 1.0	19.03 $\pm$ 6.9	0.32 $\pm$ 0.1	20.86
+ RDumb (NeurIPS’23)	49.47 $\pm$ 0.8	39.42 $\pm$ 1.5	9.77 $\pm$ 1.8	32.89
+ COME (ICLR’25)	6.34 $\pm$ 4.3	0.56 $\pm$ 0.1	0.14 $\pm$ 0.0	2.35
+ ReservoirTTA (NeurIPS’25)	48.86 $\pm$ 0.6	42.03 $\pm$ 0.9	1.15 $\pm$ 0.8	30.68
+ ASR (Ours)	51.20 $\pm$ 0.8	41.88 $\pm$ 1.6	17.10 $\pm$ 2.1	36.73
EATA (ICML’22)	49.52 $\pm$ 0.9	39.19 $\pm$ 1.7	0.82 $\pm$ 0.4	29.84

Table 1: Evaluation of our reset via a comparison with add-on methods across the commonly applicable TTA baselines using accuracy (%) on the CCC benchmark.

Table 5

	CCC			CIN-C		IN-C		IN-D109
Method	Easy	Medium	Hard	i.i.d.	non-i.i.d.	Visit 1 / 20	Mean	Visit 1 / 20	Mean
Source	33.89 $\pm$ 0.2	16.87 $\pm$ 0.2	1.27 $\pm$ 0.0	18.01 $\pm$ 0.0	18.01 $\pm$ 0.0	3.08 / 3.08	3.08 $\pm$ 0.0	32.52 / 32.52	32.52 $\pm$ 0.0
ETA (ICML’22)	43.24 $\pm$ 1.0	19.03 $\pm$ 6.9	0.32 $\pm$ 0.1	43.61 $\pm$ 0.4	43.63 $\pm$ 0.4	30.64 / 35.80	35.88 $\pm$ 1.2	41.24 / 34.21	37.22 $\pm$ 2.1
RoTTA (CVPR’23)	2.28 $\pm$ 0.6	1.76 $\pm$ 0.6	0.69 $\pm$ 0.2	29.05 $\pm$ 2.0	29.71 $\pm$ 1.7	12.45 / 12.96	17.60 $\pm$ 2.8	39.89 / 34.34	40.61 $\pm$ 3.1
ViDA (ICLR’24)	12.68 $\pm$ 0.8	5.75 $\pm$ 0.5	0.42 $\pm$ 0.0	17.76 $\pm$ 0.1	17.76 $\pm$ 0.1	3.09 / 2.84	2.99 $\pm$ 0.1	0.01 / 0.01	0.01 $\pm$ 0.0
PALM (AAAI’25)	1.56 $\pm$ 0.2	0.74 $\pm$ 0.3	0.13 $\pm$ 0.0	12.69 $\pm$ 6.3	12.08 $\pm$ 6.1	24.66 / 30.98	30.70 $\pm$ 1.4	13.86 / 1.42	2.06 $\pm$ 2.7

Table 4: Comparison with state-of-the-art continual TTA approaches across four datasets using accuracy (%). Results for each level of CCC (Easy / Medium / Hard) are averaged over nine variations, considering three different corruption orderings and three corruption evolving speeds. CIN-C results are averaged over ten runs. In the non-i.i.d. setting, we use a Dirichlet parameter

\delta=0.1

, following prior works ( Gong et al., 2022 ; Yuan et al., 2023 ) . For IN-C and IN-D109, we report averages at the first and last (20th) visits, as well as overall averages across all visits. Gray denotes model collapse, indicating performance worse than that of the source model ( Press et al., 2023 ) .

$\mathcal{C}_{t}$ ( 1 )	$r_{t}$ ( 3 )	$\bar{\mathcal{F}}$ ( 4 )	$\lambda_{0}$ ( 6 )	$\mu_{0}$ ( 7 )	Easy	Medium	Hard	Mean
✗	✗	✗	✗	✗	49.74	40.19	11.81	33.91
✗	✓	✓	✓	✗	49.83	40.58	17.16	35.86
✓	✗	✓	✓	✓	49.83	40.35	15.99	35.39
✓	✓	✗	✓	✓	51.04	42.19	20.18	37.80
✓	✓	✓	✗	✓	51.07	42.33	20.27	37.89
✓	✓	✓	✓	✗	50.82	41.86	20.70	37.79

Table 5: Effect of components in ASR on ROID .

Figure 7: Comparison of ETA / ROID and their variants with RDumb and ASR over different Dirichlet parameters

\delta

on non-i.i.d. CIN-C. The lower the

\delta

, the more imbalanced the label distribution.

Table 10

Figure 8: Performance comparison over revisits on IN-C with CDC settings. Domains dynamically change over time, controlled by

\delta

, and are color-coded to visualize their dynamics.

Figure 9: Histogram of predictions on CCC-Hard for randomly selected ten fixed class labels, comparing ETA, RDumb, and ASR to evaluate robustness against model collapse. Results are measured every

10^{3}

batches, with class labels color-coded consistently.

Appendix figures & tables34 assets

Supplementary material from the paper’s appendix.

Appendix

Excluded logits	Pearson correlation
None	0.88
Top-1	0.85
Top-10%	0.77

Appendix

Table 8: Correlation between

\mathcal{C}_{t}

and accuracy after eliminating large-magnitude logits.

Revisit #	$\varepsilon$	1	5	10	15	20	Mean
ETA	-	30.64	36.76	36.16	35.96	35.80	35.88
+ w/ delay	0.001	28.42	33.39	37.80	38.82	39.02	36.25
+ w/ delay	0.01	27.94	28.06	27.94	28.30	28.12	28.29
+ ASR (Ours)	-	28.68	33.00	38.60	38.97	39.10	36.90
ROID	-	35.32	38.00	38.08	38.15	38.02	37.96
+ w/ delay	0.001	35.08	40.01	41.42	41.96	41.54	40.85

Appendix

Table 9: Accuracy (%) under the standard and delaying reset settings on IN-C. Here,

\varepsilon

denotes the threshold for delaying a reset; a larger

\varepsilon

value encourages stronger reset delaying.

Method	# Trainable	# Param	Time	Acc.
ETA	53.1K	25.5M	.083	21.72
ROID	53.1K	51.1M	.125	33.91
+ RDumb	53.1K	51.1M	.125	35.39
+ ASR (Ours)	53.1K	51.2M	.200	38.32
+ w/o recovery (Sec. 3.4 )	53.1K	51.1M	.200	37.80
+ w/o on-the-fly (Sec. 3.5 )	53.1K	51.2M	.181	37.89

Appendix

Table 10: Computational efficiency on CCC. # Trainable indicates the number of learnable parameters; # Param the total number of parameters; Time the seconds per batch (64 samples); and Acc. the average accuracy (%) across all CCC levels.

Revisit #	…	10	15	20	Mean
ETA + ASR (Ours)	…	+0.58	+0.08	+0.02	+0.24
+ w/o knowledge recovery	…	-1.94	-1.16	-0.76	-0.56
ROID + ASR (Ours)	…	+0.16	+0.24	+0.01	+0.12
+ w/o knowledge recovery	…	-0.52	-0.42	-0.10	-0.14

Appendix

Table 11: Knowledge recovery measured across multiple revisits on IN-C.

Revisit #	1	…	10	15	20	Mean
ETA	30.64	…	36.16	35.96	35.80	35.88
+ ASR (Ours)	28.68	…	38.60	38.97	39.10	36.90
+ w/o knowledge recovery	28.64	…	37.45	36.49	36.34	36.56
ROID	35.32	…	38.08	38.15	38.02	37.96
+ ASR (Ours)	35.66	…	42.20	42.06	42.96	41.56
+ w/o knowledge recovery	35.35	…	41.64	41.64	41.19	40.96

Appendix

Table 12: Accuracy (%) comparison across multiple revisits on IN-C.

Split #	1	2	3	4	5	6	7	8	9
ROID	12.64	15.79	13.28	9.63	12.65	11.81	10.66	8.72	17.12
+ ASR (Ours)	20.99	22.51	20.40	21.22	22.93	21.36	22.37	23.84	24.25
+ w/o recovery	20.95	22.50	20.35	18.28	22.69	20.18	9.70	15.67	23.57

Appendix

Table 13: Accuracy (%) comparison across 9 splits in CCC-Hard.

Figure 12: Hyperparameter sensitivity analysis showing robustness to variations around the default.

Hyperparameter	Reference	ResNet-50	ViT-B-16
$\alpha_{0}$ : Initialization factor for $\bar{\mathcal{C}}_{t-1}$	Below Eq. ( 2 ) in Sec. 3.3	0.5	$5.0\times 10^{-4}$
$\mu_{\mathcal{C}}$ : EMA update momentum for $\bar{\mathcal{C}}_{t-1}$	Eq. ( 2 ) in Sec. 3.3	0.995	0.995
$r_{0}$ : Minimum reset proportion	Eq. ( 3 ) in Sec. 3.3	0.5	0.5
$\lambda_{r}$ : Reset proportion scaling factor	Eq. ( 3 ) in Sec. 3.3	20.0	0.1
$\lambda_{\mathcal{F}}$ : Fisher regularization coefficient	Eq. ( 4 ) in Sec. 3.4	5.0	5.0
$\lambda_{0}$ : Initialization factor for $\lambda_{\mathcal{F}}$	Eq. ( 6 ) in Sec. 3.5	5.0	5.0

Appendix

Table 14: Hyperparameter settings for ResNet-50 and ViT-B-16 used across all benchmarks.

Figure 13: Prediction concentration in Eq. ( 1 ) over time under 15 corruption types in CIN-C. Dashed vertical lines ( red ) indicate corruption boundaries. Colored ellipses highlight abrupt dynamics associated with domain shifts ( yellow : abrupt decline; green : abrupt rise).

Level	Corruption Types
Easy	Gaussian_noise, Shot_noise, Impulse_noise, Contrast
Medium	Gaussian_noise, Shot_noise, Impulse_noise, Defocus_blur,
	Glass_blur, Motion_blur, Zoom_blur, Snow, Frost, Fog,
	Contrast, Elastic, Pixelate
Hard	Gaussian_noise, Shot_noise, Impulse_noise, Defocus_blur,
	Glass_blur, Motion_blur, Zoom_blur, Snow, Frost, Fog,

Appendix

Table 15: Corruption types in a transition path for each adaptation difficulty.

Reset interval	Easy	Medium	Hard	Mean
Fixed
$T$ = 1000	15.96	5.91	1.10	7.66
$T$ = 10000	49.88	39.69	15.75	35.11
$T$ = 20000	49.83	40.58	17.16	35.86
$T$ = 50000	49.90	40.16	14.26	34.77
Dynamic

Appendix

Table 16: Comparison with our variants using fixed reset intervals

T

in terms of accuracy (%).

Reset target	Easy	Medium	Hard	Mean
Fixed
20%	49.27	40.01	16.83	35.37
50%	50.91	42.07	20.80	37.93
80%	50.72	41.40	19.68	37.27
100%	49.83	40.35	15.99	35.39
Dynamic

Appendix

Table 17: Comparison with our variants resetting a fixed % of layers closer to the output.

Method	Easy	Medium	Hard	Mean
ROID	49.74	40.19	11.81	33.91
+ SAR	49.73	40.06	5.29	31.69
+ RDumb	49.56	39.77	14.06	34.46
+ DA-TTA	45.98	35.76	15.53	32.42
+ ASR (Ours)	50.70	41.72	19.36	37.26

Appendix

Table 18: Accuracy (%) comparison across various reset schemes on CCC.

Method	Easy	Medium	Hard	Mean
CMA-only	50.03	41.13	18.42	36.53
Hybrid (Ours)	51.19	42.42	21.36	38.32

Appendix

Table 19: Hybrid knowledge accumulation (Ours) vs. CMA-only baseline.

$\lambda_{\mathcal{F}}$	Range	Easy	Medium	Hard	Mean
$\lambda_{0}$	$\{\lambda_{0}\}$	51.07	42.33	20.27	37.89
$\phi_{t}$	$[0,1]$	51.09	42.26	20.56	37.97
$\lambda_{0}\cdot(1-\phi_{t})^{2}$	$[\lambda_{0},0]$	51.15	42.34	20.42	37.97
$\lambda_{0}\cdot\phi_{t}$	$[0,\lambda_{0}]$	51.16	42.40	21.27	38.28
$\lambda_{0}\cdot\phi_{t}^{2}$	$[0,\lambda_{0}]$	51.19	42.42	21.36	38.32

Appendix

Table 20: Comparison with different expressions for

\lambda_{\mathcal{F}}

across all CCC levels using accuracy (%).

$\mu_{\mathcal{C}}$	Range	Easy	Medium	Hard	Mean
$1-\mu_{0}$	$\{1-\mu_{0}\}$	50.82	41.86	20.70	37.79
$\phi_{t}$	$[0,1]$	51.48	42.42	0.31	31.40
$1-\mu_{0}\cdot\phi_{t}$	$[1,1-\mu_{0}]$	51.17	42.47	4.07	32.57
$1-\mu_{0}\cdot(1-\phi_{t}^{2})$	$[1-\mu_{0},1]$	51.14	42.40	21.11	38.22
$1-\mu_{0}\cdot(1-\phi_{t})$	$[1-\mu_{0},1]$	51.19	42.42	21.36	38.32

Appendix

Table 21: Comparison with different expressions for

\mu_{\mathcal{C}}

across all CCC levels using accuracy (%).

	Recurring visit $\xrightarrow{\hskip 184.9429pt}$
$\delta=1.0$	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	Mean
ETA	30.62	35.77	36.19	36.21	36.14	35.99	35.84	35.76	35.73	35.63	35.42	35.43	35.31	35.33	35.30	35.22	35.25	35.20	35.17	35.08	35.33
+ RDumb	30.62	35.73	36.20	30.88	35.24	36.24	30.83	34.62	36.11	32.01	35.14	36.39	33.46	34.56	36.67	34.33	32.97	36.23	36.16	31.11	34.28
+ ASR (Ours)	28.48	30.25	33.48	33.03	32.42	35.28	35.33	35.27	35.72	36.19	36.85	37.43	38.49	38.34	38.65	38.41	38.94	38.92	38.96	38.98	35.97
ROID	34.11	36.82	37.04	36.99	36.86	36.78	37.20	37.05	36.59	36.89	36.68	36.74	36.86	36.95	36.66	36.83	36.62	36.68	36.67	36.84	36.69
+ RDumb	34.28	36.93	36.91	34.00	36.74	36.85	34.27	37.04	37.14	34.02	36.41	36.76	34.37	36.37	36.80	34.67	35.94	36.63	36.43	34.23	35.84

Appendix

Table 22: Results on IN-C with CDC for

\delta=1.0

across revisit steps.

	Recurring visit $\xrightarrow{\hskip 184.9429pt}$
$\delta=10.0$	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	Mean
ETA	29.79	34.80	35.72	35.68	35.72	35.56	35.33	35.47	35.31	35.23	35.23	35.09	35.02	35.01	34.88	34.78	34.83	34.75	34.70	34.62	34.88
+ RDumb	29.53	35.62	36.09	31.11	35.54	36.10	30.86	35.04	36.16	32.19	34.86	36.30	32.64	34.03	36.51	34.39	33.35	36.66	36.86	29.56	34.17
+ ASR (Ours)	29.40	35.06	35.55	36.01	36.63	36.71	36.92	37.25	37.31	37.30	37.41	37.53	37.79	37.79	37.92	37.96	38.08	38.14	38.19	38.11	36.85
ROID	33.60	36.70	36.51	37.06	36.74	36.99	37.11	36.82	36.79	37.04	36.71	37.23	36.86	37.06	36.99	36.77	36.59	36.82	36.87	36.99	36.71
+ RDumb	33.76	36.66	36.67	34.30	37.01	37.26	33.51	36.69	36.76	34.44	36.16	37.19	34.38	36.55	36.52	35.05	35.77	36.98	36.34	33.80	35.79

Appendix

Table 23: Results on IN-C with CDC for

\delta=10.0

across revisit steps.

Batch size	2	4
ROID	0.13	16.91
+ RDumb	5.87	17.85
+ ASR (Ours)	6.46	25.58

Appendix

Table 24: Accuracy (%) under truly small batch sizes on CCC-Easy.

Level	Corruption Types
Easy	Motion_blur, Snow, Fog, Elastic, JPEG
Medium	Defocus_blur, Glass_blur, Zoom_blur,
	Frost, Contrast, Pixelate
Hard	Gaussian_noise, Shot_noise, Impulse_noise

Appendix

Table 25: Corruption types for each level of CIFAR10-C.

Level	Corruption Types
Easy	Impulse_noise, Defocus_blur, Motion_blur,
	Zoom_blur, Snow, Brightness, Elastic
Medium	Glass_blur, Frost, Fog, Contrast, JPEG
Hard	Gaussian_noise, Shot_noise, Pixelate

Appendix

Table 26: Corruption types for each level of CIFAR100-C.

Figure 14: Comparison of ETA / ROID and their variants combined with RDumb or ASR across three levels of CIFAR10-C using accuracy (%), averaged over 1000 recurring visits of corruptions.

Figure 15: Comparison of ETA / ROID and their variants combined with RDumb or ASR across three levels of CIFAR100-C using accuracy (%), averaged over 1000 recurring visits of corruptions.

ViT-Tiny	CCC-Hard	CCC-Medium	CCC-Easy
ETA	2.29	34.20	47.09
+ RDumb	4.45	32.51	45.51
+ ASR (Ours)	5.30	36.48	47.23
ROID	0.10	32.14	45.29
+ RDumb	0.10	31.61	44.92
+ ASR (Ours)	0.10	34.47	45.64

Appendix

Table 27: Accuracy (%) comparison on ViT-Tiny.

Transition speed	1000			2000			5000			Acc. (%)
Corruption ordering	43	44	45	43	44	45	43	44	45	Mean
Source	33.89	33.97	33.95	33.69	33.90	33.99	33.34	34.06	34.23	33.89 $\pm$ 0.2
RMT (CVPR’23)	48.15	46.78	47.38	46.70	46.44	47.80	48.61	45.03	48.07	47.22 $\pm$ 1.0
RoTTA (CVPR’23)	1.76	1.49	1.88	2.43	1.84	2.28	2.50	3.06	3.28	2.28 $\pm$ 0.6
SANTA (TMLR’23)	47.33	47.47	47.49	47.87	47.68	47.77	48.32	47.11	48.10	47.68 $\pm$ 0.4
LAW (WACV’24)	2.71	2.50	2.92	2.99	2.44	3.25	2.79	2.20	3.55	2.82 $\pm$ 0.4

Appendix

Table 28: Performance comparison with state-of-the-art methods on CCC-Easy across nine variants: three corruption transition speeds (1000 / 2000 / 5000) and three corruption orderings that are determined by random seeds (43 / 44 / 45).

Transition speed	1000			2000			5000			Acc. (%)
Corruption ordering	43	44	45	43	44	45	43	44	45	Mean
Source	16.95	16.78	16.95	16.59	16.87	16.97	16.57	16.91	17.20	16.87 $\pm$ 0.2
RMT (CVPR’23)	35.48	35.38	35.60	36.07	35.65	34.08	35.41	31.42	36.09	35.02 $\pm$ 1.4
RoTTA (CVPR’23)	1.23	1.00	1.36	1.84	1.31	1.70	2.76	2.08	2.54	1.76 $\pm$ 0.6
SANTA (TMLR’23)	33.75	33.77	34.17	35.65	34.18	34.26	35.94	33.79	34.57	34.45 $\pm$ 0.8
LAW (WACV’24)	1.56	1.09	1.50	1.66	0.64	1.57	1.38	0.76	1.57	1.30 $\pm$ 0.4

Appendix

Table 29: Performance comparison with state-of-the-art methods on CCC-Medium across nine variants: three corruption transition speeds (1000 / 2000 / 5000) and three corruption orderings that are determined by random seeds (43 / 44 / 45).

Transition speed	1000			2000			5000			Acc. (%)
Corruption ordering	43	44	45	43	44	45	43	44	45	Mean
Source	1.29	1.23	1.31	1.31	1.23	1.30	1.33	1.19	1.25	1.27 $\pm$ 0.0
RMT (CVPR’23)	12.13	13.18	13.12	9.43	10.74	12.83	7.73	0.86	9.27	9.92 $\pm$ 3.7
RoTTA (CVPR’23)	0.50	0.77	0.74	0.66	0.77	0.96	0.79	0.17	0.87	0.69 $\pm$ 0.2
SANTA (TMLR’23)	9.28	9.93	9.16	9.08	9.89	9.14	8.96	9.77	9.97	9.46 $\pm$ 0.4
LAW (WACV’24)	0.34	0.17	0.22	0.31	0.17	0.22	0.27	0.16	0.20	0.23 $\pm$ 0.1

Appendix

Table 30: Performance comparison with state-of-the-art methods on CCC-Hard across nine variants: three corruption transition speeds (1000 / 2000 / 5000) and three corruption orderings that are determined by random seeds (43 / 44 / 45).

Method	1	2	3	4	5	6	7	8	9	10	Mean
Source	18.01	18.01	18.01	18.01	18.01	18.01	18.01	18.01	18.01	18.01	18.01 $\pm$ 0.0
RMT (CVPR’23)	47.68	45.48	44.09	43.87	46.48	45.70	44.68	44.86	43.61	43.64	45.01 $\pm$ 1.3
+ Source-free	42.33	39.13	33.49	36.63	40.32	38.21	37.75	34.24	33.02	33.63	36.88 $\pm$ 3.1
RoTTA (CVPR’23)	27.21	31.85	27.23	24.99	28.56	30.99	30.55	29.61	30.70	28.85	29.05 $\pm$ 2.0
SANTA (TMLR’23)	40.00	39.85	39.83	39.77	39.84	39.53	39.83	39.85	39.63	39.98	39.81 $\pm$ 0.1
LAW (WACV’24)	22.91	17.65	1.14	14.63	24.72	17.70	10.91	11.27	11.66	2.06	13.47 $\pm$ 7.4

Appendix

Table 31: Accuracy (%) on CIN-C over ten random permutations of the corruption order.

Method	1	2	3	4	5	6	7	8	9	10	Mean
Source	18.01	18.01	18.01	18.01	18.01	18.01	18.01	18.01	18.01	18.01	18.01 $\pm$ 0.0
RMT (CVPR’23)	46.53	44.99	42.80	44.17	45.95	44.01	43.88	43.59	42.99	42.92	44.18 $\pm$ 1.2
+ Source-free	42.07	38.68	32.79	36.20	40.16	37.47	37.30	34.04	32.34	33.08	36.41 $\pm$ 3.2
RoTTA (CVPR’23)	29.33	32.30	27.21	27.16	29.09	32.17	30.24	30.19	30.41	28.96	29.71 $\pm$ 1.7
SANTA (TMLR’23)	39.60	39.38	39.28	39.28	39.54	39.52	39.37	39.44	39.42	39.16	39.40 $\pm$ 0.1
LAW (WACV’24)	21.82	15.81	1.47	15.34	24.20	16.41	9.00	13.86	13.06	2.74	13.37 $\pm$ 6.9

Appendix

Table 32: Accuracy (%) on non-i.i.d. CIN-C over ten random permutations of the corruption order.

	Recurring visit $\xrightarrow{\hskip 184.9429pt}$
Method	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	Mean
Source	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08	3.08 $\pm$ 0.0
RMT	27.63	33.91	37.28	39.08	39.99	40.78	41.18	41.36	41.72	41.76	41.76	41.92	42.01	42.02	41.96	42.05	42.10	42.08	42.08	42.09	40.24 $\pm$ 3.5
+ Source-free	27.63	34.15	37.14	38.11	38.70	38.87	39.22	39.43	39.40	39.60	39.60	39.60	39.71	39.80	39.76	39.68	39.74	39.79	39.76	39.76	38.47 $\pm$ 2.8
RoTTA	12.45	17.22	19.19	20.77	19.92	21.29	21.88	21.23	19.71	19.25	18.70	18.00	17.33	16.91	16.20	15.58	14.97	14.44	13.97	12.96	17.60 $\pm$ 2.8
SANTA	27.28	27.75	27.30	27.20	27.10	26.94	27.04	26.81	26.91	26.59	26.49	26.42	26.53	26.39	26.25	26.38	26.18	26.17	26.25	25.94	26.70 $\pm$ 0.5

Appendix

Table 33: Accuracy (%) on IN-C across 20 recurring visits of the domain sequence.

	Recurring visit $\xrightarrow{\hskip 184.9429pt}$
Method	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	Mean
Source	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52	32.52 $\pm$ 0.0
RMT	43.16	45.54	46.26	46.51	46.74	46.78	46.81	46.83	46.88	46.88	46.87	46.88	46.90	46.89	46.90	46.90	46.89	46.88	46.86	46.87	46.56 $\pm$ 0.8
+ Source-free	42.24	43.73	43.82	43.92	43.94	43.98	43.81	43.86	43.87	43.88	43.77	43.81	43.80	43.76	43.75	43.71	43.68	43.68	43.68	43.65	43.72 $\pm$ 0.4
RoTTA	39.89	43.06	44.03	44.36	44.34	43.94	43.66	43.26	42.71	42.07	41.41	40.63	40.04	39.38	38.66	37.85	37.04	36.20	35.23	34.34	40.61 $\pm$ 3.1
SANTA	41.52	41.68	41.74	41.66	41.75	41.80	41.68	41.59	41.65	41.54	41.44	41.47	41.39	41.58	41.50	41.41	41.42	41.34	41.40	41.29	41.54 $\pm$ 0.1

Appendix

Table 34: Accuracy (%) on IN-D109 across 20 recurring visits of the domain sequence.

Transition speed	1000			2000			5000			Acc. (%)
Corruption ordering	43	44	45	43	44	45	43	44	45	Mean
Source	54.74	55.18	54.47	54.97	55.03	54.77	55.09	54.88	55.12	54.92 $\pm$ 0.2
CMF (ICLR’24)	60.68	60.96	60.59	61.51	61.74	61.51	62.66	62.52	61.52	61.52 $\pm$ 0.7
CMAE (CVPR’24)	48.11	49.94	48.19	50.68	50.44	52.57	51.04	54.89	54.50	51.15 $\pm$ 2.3
REM (ICML’25)	65.82	66.12	65.79	66.14	66.23	66.05	66.93	66.02	66.36	66.16 $\pm$ 0.3
ETA (ICML’22)	47.95	47.68	47.47	48.71	15.74	48.44	49.92	49.78	49.93	45.07 $\pm$ 10.4

Appendix

Table 35: Performance comparison with state-of-the-art methods on CCC-Easy across nine variants: three corruption transition speeds (1000 / 2000 / 5000) and three corruption orderings that are determined by random seeds (43 / 44 / 45).

Transition speed	1000			2000			5000			Acc. (%)
Corruption ordering	43	44	45	43	44	45	43	44	45	Mean
Source	41.76	41.47	40.49	42.32	41.10	41.49	42.15	42.39	42.51	41.74 $\pm$ 0.6
CMF (ICLR’24)	52.17	51.63	51.87	34.21	53.49	53.59	55.41	55.33	55.82	51.50 $\pm$ 6.3
CMAE (CVPR’24)	41.76	36.63	41.88	43.16	40.52	44.40	45.02	46.41	51.52	43.48 $\pm$ 3.9
REM (ICML’25)	57.34	57.23	56.92	58.36	57.28	57.68	58.94	58.41	59.71	57.99 $\pm$ 0.9
ETA (ICML’22)	34.62	23.74	28.04	34.04	34.54	36.05	35.36	38.07	38.97	33.71 $\pm$ 4.6

Appendix

Table 36: Performance comparison with state-of-the-art methods on CCC-Medium across nine variants: three corruption transition speeds (1000 / 2000 / 5000) and three corruption orderings that are determined by random seeds (43 / 44 / 45).

Transition speed	1000			2000			5000			Acc. (%)
Corruption ordering	43	44	45	43	44	45	43	44	45	Mean
Source	14.40	15.44	15.40	14.16	15.38	14.10	13.90	15.31	15.40	14.83 $\pm$ 0.6
CMF (ICLR’24)	1.22	0.30	2.74	2.17	0.14	0.85	3.23	0.13	5.34	1.79 $\pm$ 1.7
CMAE (CVPR’24)	26.47	26.78	22.70	25.95	28.33	24.96	26.60	30.27	30.20	26.92 $\pm$ 2.3
REM (ICML’25)	3.80	8.53	7.03	5.58	5.94	9.39	10.67	38.45	9.31	10.97 $\pm$ 9.9
ETA (ICML’22)	1.34	0.33	1.66	0.99	1.34	1.02	1.97	1.58	0.79	1.22 $\pm$ 0.5

Appendix

Table 37: Performance comparison with state-of-the-art methods on CCC-Hard across nine variants: three corruption transition speeds (1000 / 2000 / 5000) and three corruption orderings that are determined by random seeds (43 / 44 / 45).

Test-Time Adaptation

Test-time adaptation updates or recalibrates pretrained models at inference to handle distribution shifts, often without target labels.

See all 322 papers

Jul 9, 2026

Continual Test-Time Adaptation in Computer Vision: Methods, Benchmarks, and Future Directions

Deep neural nets achieve remarkable performance when training and test data share the same distribution, but this assumption frequently breaks in real-world deployment, where data undergoes continual distributional shifts. Continual Test-Time Adaptation (CTTA) addresses this challenge by adapting pretrained models to non-stationary target distributions on-the-fly, without access to source data or labeled targets, while mitigating two critical failure modes: catastrophic forgetting of source knowledge and error accumulation from noisy pseudo-labels over extended time horizons. In this comprehensive survey, we formally define the CTTA problem, analyze the diverse continual domain shift patterns that characterize different evaluation protocols, and propose a hierarchical taxonomy that categorizes existing methods into three families: optimization-based strategies (entropy minimization, pseudo-labeling, parameter restoration), parameter-efficient methods (normalization layer adaptation, adaptive parameter selection), and architecture-based approaches (teacher-student frameworks, adapters, visual prompting, masked modeling). We systematically review representative methods within each category and present comparative benchmarks and experimental results across standard evaluation settings. Finally, we discuss the limitations of current approaches and highlight emerging research directions, including the adaptation of foundation models and black-box systems, thereby providing a roadmap for future research in robust continual test-time adaptation.

Sarthak Kumar Maharana, Shambhavi Mishra, Yunbei Zhang +6

The University of Texas at Dallas, USA · LIVIA ETS Montreal, ILLS International Laboratory on Learning Systems (ILLS), McGill - ÉTS - MILA - CNRS - Université Paris-Saclay - CentraleSupélec · Tulane University, USA +2

Source-Free Domain Adaptation Distribution Shift Robustness Test-Time Adaptation Continual Test-Time Adaptation

Ask PDF

Aug 23, 2026

When Test-Time Adaptation Helps, Harms, or Becomes Inactive: A Condition-Level Study on CIFAR-10-C

Test-time adaptation (TTA) aims to improve model robustness under distribution shift by adapting a source model using unlabeled test data. Although methods such as TENT and EATA have demonstrated gains on corrupted data, aggregate accuracy can obscure the conditions under which adaptation fails or provides little benefit. We present a controlled comparison of three TTA strategies---BatchNorm-statistics adaptation (BN-Adapt), entropy-minimization adaptation (TENT), and reliability-filtered adaptation (a scoped re-implementation of EATA)---against an unadapted source model on the full CIFAR-10-C benchmark, covering 15 corruption types and 5 severity levels. All three methods improve mean accuracy over the source model by 12.2--13.3 percentage points (Wilcoxon signed-rank

p < 10^{-12}

). However, each method underperforms the source model on 8.0--9.3% of conditions, with failures concentrated in low-severity corruptions where the source model already performs near ceiling, particularly brightness, fog, contrast, and defocus blur. We further find that EATA closely tracks the gradient-free BN-Adapt baseline, with a mean absolute difference of 0.09 percentage points, compared with 1.08 percentage points relative to TENT. This suggests that reliability filtering can substantially restrict effective adaptation, causing EATA to behave more like a BatchNorm-statistics baseline than an entropy-minimization method. These results show that aggregate accuracy alone can mask systematic TTA failure modes and motivate condition-level evaluation of when adaptation helps, harms, or becomes effectively inactive.

Sreeja Guha Majumdar, Aratrika Saha

Heritage Institute of Technology Kolkata, West Bengal, India

Test-Time Adaptation Corruption Robustness

Ask PDF

Apr 7, 2025

TestDG: Test-time Domain Generalization for Continual Test-time Adaptation

This paper studies continual test-time adaptation (CTTA), the task of adapting a model to constantly changing unseen domains in testing while preserving previously learned knowledge. Existing CTTA methods mostly focus on adaptation to the current test domain only, overlooking generalization to arbitrary test domains a model may face in the future. To tackle this limitation, we present a novel online test-time domain generalization framework for CTTA, dubbed TestDG. TestDG aims to learn features invariant to both current and previous test domains on the fly during testing, improving the potential for effective generalization to future domains. To this end, we propose a new model architecture and a test-time adaptation strategy dedicated to learning domain-invariant features, along with prototype selection and update mechanisms for effectively managing information from previous test domains. TestDG achieved state of the art on three of four public CTTA benchmarks and performed on par with the best prior method on ImageNet-C. Moreover, it showed superior generalization to unseen test domains.

Sohyun Lee, Nayeong Kim, Juwon Kang +2

POSTECH · GenGenAI · KAIST AI

Continual Test-Time Adaptation Cross-Domain Generalization Invariant Representation Learning

Ask PDF

Abstract

Test-Time Adaptation

Explore similar work

Continual Test-Time Adaptation in Computer Vision: Methods, Benchmarks, and Future Directions

When Test-Time Adaptation Helps, Harms, or Becomes Inactive: A Condition-Level Study on CIFAR-10-C

TestDG: Test-time Domain Generalization for Continual Test-time Adaptation