The Batch Size Can Be One: Test-Time Adaptation from Classifier Geometry

Jaeho Hwang; Yujin Oh; Juhyeon Shin; Juhye Lee; Sungroh Yoon; Uiwon Hwang

cs.LGOct 27, 2026

The Batch Size Can Be One: Test-Time Adaptation from Classifier Geometry

Authors: Jaeho Hwang, Yujin Oh, Juhyeon Shin, Juhye Lee, Sungroh Yoon, Uiwon Hwang

Organizations: Human-Centered Artificial Intelligence Research Institute, Ewha Womans University · Division of Artificial Intelligence & Software, Ewha Womans Univeristy · Interdisciplinary Program in Artificial Intelligence, Seoul National University

Abstract

Test-time adaptation updates a deployed model on an unlabeled test stream. In practice this stream often arrives one sample at a time, with the classes in no particular order. A single sample then provides limited information for estimating distribution-level statistics, and selecting samples by confidence can admit predictions that are wrong yet confident. A corruption, however, shifts the features while leaving the decision regions of the frozen classifier unchanged. We therefore propose Origin-recentered Test-Time Adaptation (OriTTA), which constructs fixed geometric targets from the source model and adapts selected features toward them. Seen from the origin of the feature space, the classes lean toward a common direction. We choose a new origin to bring class scores closer together while keeping it near the mean feature of one synthetic prototype per class. From this origin, prototype directions spread more evenly, and each class obtains an axis and an ideal point before any test data arrives. During adaptation, OriTTA aligns selected features with their predicted class axis and penalizes excessive radius, keeps one input per predicted class in memory to limit feature changes for classes absent from the stream, and resets the model when re-classifying this memory reveals collapse. Without tuning for each setting, OriTTA achieves the highest mean accuracy at batch size 1 on CIFAR-10-C, CIFAR-100-C and ImageNet-C in all sixteen combinations of four backbones, two protocols and two stream orders. On ImageNet-C fog, recent methods collapse in up to two thirds of one hundred runs, while OriTTA never does.

Figures & tables

Figure 1: Overview of OriTTA. (a) The frozen classifier and synthetic anchors define a fixed reference frame. (b) Reliable projected features undergo angular alignment and radial regularization. (c) Stored images are re-encoded for alignment to fixed ideal points, while concentrated current predictions over the memory trigger resets. Snowflakes denote frozen quantities.

episodic	gaus	shot	impu	defo	glas	moti	zoom	snow	fros	fog	brit	cont	elas	pixe	jpeg	Average
TENT	12.35	20.22	13.91	13.16	6.46	18.07	18.23	16.00	20.65	1.48	69.80	42.04	6.22	49.38	53.75	24.11 $\pm$ 0.40
EATA	29.49	31.55	30.23	17.93	19.32	30.45	32.62	42.73	41.21	45.55	69.40	43.66	32.98	52.13	54.02	38.22 $\pm$ 0.21
SAR	28.63	31.01	29.40	18.46	15.23	28.82	30.43	44.75	44.39	33.60	71.95	44.61	13.03	47.83	56.24	35.89 $\pm$ 0.11
DeYO	40.87	43.33	42.08	22.07	23.69	38.63	28.12	52.26	50.21	18.88	73.19	51.27	42.97	57.31	58.54	42.89 $\pm$ 1.95
ReCAP	42.07	44.35	43.17	24.69	25.05	41.73	39.64	54.67	52.16	46.54	73.33	53.52	41.54	60.27	60.20	46.86 $\pm$ 1.52
SaTeen	43.20	45.44	44.25	25.80	26.25	42.44	26.68	55.29	52.69	19.30	73.34	54.00	48.71	60.77	60.13	45.22 $\pm$ 2.20

Table 1: IN-RN50 under the four protocols. Per-corruption standard deviations are in Appendix E . The best result in each column is shown in bold and the second best is underlined.

	zoom	elastic	fog	collapse
TENT	18.2 $\pm$ 0.6	6.1 $\pm$ 0.5	1.5 $\pm$ 0.1	100%
EATA	32.9 $\pm$ 1.6	33.5 $\pm$ 1.9	46.2 $\pm$ 2.1	0%
SAR	30.4 $\pm$ 0.3	14.1 $\pm$ 1.4	33.2 $\pm$ 2.2	0%
DeYO	33.1 $\pm$ 6.8	43.0 $\pm$ 0.5	21.0 $\pm$ 25.2	66%
ReCAP	40.4 $\pm$ 3.6	45.1 $\pm$ 8.2	39.2 $\pm$ 25.4	33%
SaTeen	27.2 $\pm$ 9.1	48.8 $\pm$ 0.4	23.6 $\pm$ 27.7	62%

Table 2: One hundred seeds (

2001

–

2100

) on the three corruptions with the least stable baselines for IN-RN50, episodic, shuffled. Collapse is the fraction of seeds in which accuracy of fog corruption falls below

5\%

C10

C100

IN-ViT

episodic

continual

episodic

continual

episodic

continual

shuffled

class-blk

shuffled

class-blk

shuffled

class-blk

shuffled

class-blk

shuffled

class-blk

shuffled

class-blk

TENT

46.70

\pm

2.77

43.52

\pm

6.84

15.66

\pm

0.77

11.04

\pm

0.78

25.83

\pm

1.73

18.18

\pm

1.18

1.24

\pm

0.11

1.17

\pm

0.12

50.95

\pm

0.13

51.14

\pm

0.23

47.85

\pm

2.96

48.85

\pm

4.09

EATA

56.66

\pm

0.08

56.63

\pm

0.05

56.96

\pm

0.23

56.94

\pm

0.17

54.49

\pm

0.27

48.98

\pm

0.44

29.22

\pm

1.37

18.05

\pm

0.68

62.18

\pm

0.07

45.63

\pm

0.10

59.90

\pm

0.13

9.06

\pm

0.31

SAR

56.33

\pm

0.03

56.96

\pm

0.10

55.99

\pm

0.87

55.49

\pm

2.39

44.58

\pm

0.24

47.10

\pm

0.56

20.36

\pm

7.07

31.24

\pm

6.53

57.79

\pm

0.32

44.58

\pm

0.42

59.20

\pm

0.10

50.41

\pm

0.92

DeYO

51.92

\pm

2.94

57.34

\pm

6.14

44.42

\pm

3.45

37.86

\pm

6.62

10.49

\pm

1.02

12.48

\pm

1.56

1.89

\pm

0.55

1.91

\pm

0.26

62.91

\pm

0.62

62.93

\pm

1.09

60.70

\pm

0.54

60.60

\pm

0.34

Table 3: Mean accuracy over the 15 corruptions on C10, C100, and IN-ViT. Full results are in Appendix E .

	C10	C100	IN-RN50	IN-ViT
episodic	10.31	1.14	0.12	0.12
continual	10.30	0.80	0.35	0.19

Table 4: CoTTA at batch size 1 under our protocol, shuffled stream.

	shift			rows dev. $\downarrow$		pairwise deviation $\downarrow$			axis $\uparrow$		in cone $\uparrow$
	$\\|\omega\\|$	$\\|\omega\\|/s$	$\\|P\omega\\|/s$	$w_{c}$	$v_{c}$	at $0$	proj.	at $\omega$	at $0$	at $\omega$	at $0$	at $\omega$
CIFAR-10-C	4.77	0.96	0.05	0.07	0.07	0.58	0.38	0.36	0.79	0.82	0.00	0.00
CIFAR-100-C	12.11	0.92	0.57	0.10	0.10	0.52	0.36	0.16	0.67	0.81	0.47	0.99
ImageNet-C, RN50	6.92	0.80	0.35	0.65	0.03	0.31	0.15	0.07	0.54	0.58	1.00	1.00
ImageNet-C, ViT	33.62	0.42	0.42	0.06	0.06	0.21	0.21	0.09	0.69	0.77	1.00	1.00

Table 5: Geometry of the source model under the three operations. Rows dev. is the deviation of the classifier rows’ mutual cosines from an equiangular arrangement,

-1/(K-1)

, before and after centering. pairwise deviation is the same quantity for the prototypes

z^{a}_{c}

, measured from the origin, after projection, and from

\omega

. Axis is the median cosine between prototype

c

and its own class axis

v_{c}

, and in cone the fraction of prototypes the ideal cone admits.

		gaus	shot	impu	defo	glas	moti	zoom	snow	fros	fog	brit	cont	elas	pixe	jpeg	Average
IN-RN50	OriTTA	41.11	43.15	42.22	33.53	31.93	42.38	46.33	54.32	53.27	59.04	72.80	53.79	49.64	59.77	60.62	49.59
	no projection	38.93	40.83	39.50	31.02	27.94	39.57	43.71	53.65	52.58	48.34	72.72	52.62	45.98	57.63	60.35	47.02
	$\Delta$	$-2.18$	$-2.32$	$-2.72$	$-2.51$	$-3.99$	$-2.81$	$-2.62$	$-0.67$	$-0.69$	$\mathbf{-10.70}$	$-0.08$	$-1.17$	$-3.66$	$-2.14$	$-0.27$	$-2.57$
C100	OriTTA	48.07	52.29	51.20	70.98	54.74	69.73	70.81	63.13	63.30	58.71	71.93	63.51	62.84	64.83	59.24	61.69
	origin at $0$	44.78	49.69	48.92	69.94	53.38	68.74	69.25	61.74	61.54	56.09	71.23	60.11	61.74	60.29	58.45	59.73
	$\Delta$	$-3.29$	$-2.60$	$-2.28$	$-1.04$	$-1.36$	$-0.99$	$-1.56$	$-1.39$	$-1.76$	$-2.62$	$-0.70$	$-3.40$	$-1.10$	$\mathbf{-4.54}$	$-0.79$	$-1.96$

Table 6: Removing one operation of the frame, episodic. Top: losses computed in the full feature space, IN-RN50 (episodic, shuffled). Bottom: origin moved from

\omega

0

and projection to

\mathcal{I}

, with

s

\mu_{c}

, and

\delta

recomputed, C100 (episodic, class-blocked).

episodic	gaus	shot	impu	defo	glas	moti	zoom	snow	fros	fog	brit	cont	elas	pixe	jpeg	Average
SAR	28.53	31.11	29.44	18.56	15.14	28.83	30.53	44.16	44.36	33.30	71.93	44.67	13.87	47.73	56.22	35.89 (0)
SAR+RT	28.54	31.15	29.42	18.56	15.13	28.83	30.54	44.15	44.32	33.31	71.93	44.67	20.47 (1)	47.71	56.17	36.33 (1)
OriTTA	40.79	42.55	41.81	33.32	31.09	41.83	45.80	53.69	52.67	57.92	72.50	53.12	48.79	58.73	60.05	48.98 (0)
continual	gaus	shot	impu	defo	glas	moti	zoom	snow	fros	fog	brit	cont	elas	pixe	jpeg	Average
RDumb	25.44	22.09 (1)	22.14 (1)	15.51 (1)	14.97	20.73 (1)	21.44 (1)	28.41 (1)	22.80 (1)	37.82	66.62 (1)	32.52 (1)	15.86 (1)	34.16	48.97 (1)	28.64 (11)
SAR	28.53	41.14	44.47	15.76	19.17	11.53 (8.5)	11.17 (7.1)	15.79 (6.1)	18.37 (36.8)	18.04 (3564)	34.59 (8)	20.35 (144)	5.68 (3740)	18.72 (33.1)	20.27 (1476)	21.57 (9025)

Table 7: SAR with its own recovery, SAR with OriTTA’s detector and reset in its place (SAR+RT), and OriTTA, on IN-RN50, class-blocked. RDumb ( Press et al., 2023 ) restores EATA to its source weights every

64{,}000

images. The subscript is the mean number of resets per run, shown only where nonzero. The total is given with the average.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Dataset	Backbone	Norm	$D$	$K$	$r$
C10	CIFAR-10-C	WideResNet-28-10	BatchNorm	640	10	9
C100	CIFAR-100-C	WideResNet-40-2	BatchNorm	128	100	99
IN-RN50	ImageNet-C	ResNet-50	GroupNorm	2048	1000	999
IN-ViT	ImageNet-C	ViT-B/16	LayerNorm	768	1000	768

Appendix

Table 8: Settings.

D

is the feature dimension,

K

the number of classes and

r=\min(K-1,D)

the dimension of the ideal subspace.

	released	regenerated
$\\|\omega\\|$	$6.93$	$6.96\pm 0.02$
$s$	$8.63$	$8.62\pm 0.02$
$\delta$	$3.45$	$3.44\pm 0.01$
shift of $\omega$	–	$0.031\pm 0.001$
shift of $\bar{z}^{a}$	–	$0.033\pm 0.001$
shift of one anchor (median)	–	$0.75\pm 0.01$

Appendix

Table 9: Frame of IN-RN50 built from the released anchors and from anchors regenerated with ten other seeds (mean and standard deviation). Distances are in units of the released

s

Setting	$\\|\omega\\|$	$\\|P\omega\\|/s$	$s$	$d_{\mathrm{nn}}$	$\delta$	$\tau$	$\kappa$	$\lambda$
C10	4.77	0.05	4.95	7.01	2.33	2.23	0.33	0.1
C100	12.11	0.57	13.21	15.09	5.02	3.63	1.09	1
IN-RN50	6.92	0.35	8.62	10.35	3.45	2.94	3.48	10
IN-ViT	33.62	0.42	79.86	87.26	29.06	8.94	3.05	10

Appendix

Table 10: Quantities derived from the source model.

s

is the prototype radius,

d_{\mathrm{nn}}

the median nearest-neighbor distance between ideal points,

\delta=d_{\mathrm{nn}}/3

the radial slack,

\tau=\sqrt{s}

the temperature of

L_{\mathrm{ang}}

\kappa=0.11\sqrt{r}

the cone slope, and

\lambda=K/100

the ridge weight in Eq. 5 .

	Value	Role
$\kappa_{0}$	$0.11$	cone slope per $\sqrt{r}$ (Eq. 9 )
$t_{\min}$	$0.5$	lower bound on the radial coordinate in the cone
trust ratio	$0.2$	quantile that defines the reliable region (Eq. 10 )
warm-up	$20$	cone samples per class before the quantile engages
$\eta$	$0.02$	rate of the quantile tracker (Appendix A.5 )
$w_{\mathrm{ang}},w_{\mathrm{align}},w_{\mathrm{rad}}$	$1,1,2$	loss weights (Eq. 15 )

Appendix

Table 11: Constants of the method, shared across all settings.

	gaus (s)	fog (s)	brit (s)	total (h)
No adaptation	304.0	301.4	300.9	1.3
TENT	715.5	657.3	662.3	2.8
EATA	816.7	1097.6	1328.5	4.5
SAR	1961.1	3438.8	3414.4	12.0
DeYO	710.6	669.4	1139.1	3.5
ReCAP	1065.9	1473.2	1778.3	6.0

Appendix

Table 12: Wall-clock time on IN-RN50, episodic, shuffled, batch size 1.

			gaus	shot	impu	defo	glas	moti	zoom	snow	fros	fog	brit	cont	elas	pixe	jpeg	Total	Runs
IN-RN50	episodic	shuffled	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	0.0	0/10
	episodic	class-blocked	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	0.0	0/10
	continual	shuffled	$\cdot$	$\cdot$	1.0	$\cdot$	1.0	$\cdot$	0.3	0.7	$\cdot$	0.1	$\cdot$	0.8	0.5	0.5	0.3	5.2	10/10
	continual	class-blocked	$\cdot$	$\cdot$	0.8	0.2	0.8	$\cdot$	0.9	0.1	$\cdot$	0.6	0.1	0.1	1.0	$\cdot$	$\cdot$	4.6	10/10
IN-ViT	episodic	shuffled	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	0.0	0/10
	episodic	class-blocked	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	$\cdot$	0.0	0/10

Appendix

Table 13: Resets per run, averaged over ten seeds, per corruption. A dot is zero. Runs is the number of the ten seeds in which the detector fired at least once. In the continual protocol the corruptions are presented from left to right.

		max $\sigma$ , runs without reset	$\sigma$ at firing (median)
IN-RN50	episodic	0.037	–
	continual	–	0.258
IN-ViT	episodic	0.022	–
	continual	–	0.293
C100	episodic	0.150	–
	continual	0.030	0.320

Appendix

Table 14: Bank share

\sigma

in runs without and with a reset, pooled over both stream orders and ten seeds. The threshold is

0.25

	gaus	shot	impu	defo	glas	moti	zoom	snow	fros	fog	brit	cont	elas	pixe	jpeg	Average
OriTTA + Adam	40.79	42.55	41.81	33.32	31.09	41.83	45.80	53.69	52.67	57.92	72.50	53.12	48.79	58.73	60.05	48.98
sd	0.17	0.24	0.23	0.29	0.38	0.22	0.13	0.20	0.15	0.36	0.11	0.29	0.35	0.21	0.09	0.10
OriTTA + SGD	43.46	44.45	43.40	34.75	34.41	45.04	48.76	55.86	53.47	55.93	71.82	55.08	54.25	62.63	60.61	50.93
sd	0.20	0.24	0.27	0.27	0.37	0.28	0.18	0.19	0.24	1.03	0.19	0.28	0.39	0.23	0.18	0.32
$\Delta$	$+2.67$	$+1.90$	$+1.59$	$+1.43$	$+3.32$	$+3.21$	$+2.96$	$+2.17$	$+0.80$	$-1.99$	$-0.68$	$+1.96$	$+5.46$	$+3.90$	$+0.56$	$+1.95$

Appendix

Table 15: Adam against SGD, IN-RN50, episodic, class-blocked, ten seeds. The second line of each method is the standard deviation over seeds.

	gaus	shot	impu	defo	glas	moti	zoom	snow	fros	fog	brit	cont	elas	pixe	jpeg	Average
OriTTA (full)	40.79	42.55	41.81	33.32	31.09	41.83	45.80	53.69	52.67	57.92	72.50	53.12	48.79	58.73	60.05	48.98
w/o $L_{\mathrm{ang}}$	26.13	27.30	25.61	22.63	14.42	28.11	29.11	45.44	48.82	35.14	69.48	39.61	25.09	35.00	53.92	35.05 (-13.92)
w/o $L_{\mathrm{rad}}$	40.00	41.44	40.87	32.06	30.05	40.08	44.61	52.98	51.84	57.48	71.70	52.03	47.44	57.64	58.92	47.94 (-1.03)
w/o $L_{\mathrm{align}}$	40.49	41.36	40.71	33.08	30.23	40.19	44.59	52.62	51.53	58.28	72.35	51.85	47.89	58.28	59.35	48.19 (-0.79)

Appendix

Table 16: Removing one loss term at a time for IN-RN50, episodic, class-blocked.

		accuracy			resets per run
	block	full	w/o $L_{\mathrm{align}}$	$\Delta$	full	w/o $L_{\mathrm{align}}$
IN-RN50	50	48.98	48.19	$-0.79$	0 (0/10)	0 (0/10)
C100	100	61.69	60.27	$-1.41$	0 (0/10)	0 (0/10)
C10	1000	68.96	60.94	$-8.02$	0.5 (4/10)	37.1 (10/10)

Appendix

Table 17: Removing

L_{\mathrm{align}}

, episodic, class-blocked. Block is the number of consecutive images of one class and resets is the mean number of detector firings per run and the number of runs, out of ten, in which it fired.

C10	gaus	shot	impu	defo	glas	moti	zoom	snow	fros	fog	brit	cont	elas	pixe	jpeg	Average
OriTTA	46.96	52.50	39.10	73.22	54.49	78.07	76.05	78.28	72.51	83.49	91.52	75.95	77.83	61.40	73.03	68.96
sd	4.01	2.63	3.11	2.79	1.76	2.19	1.99	1.83	1.31	1.30	0.76	1.52	1.48	3.99	2.02	1.51
w/o $L_{\mathrm{align}}$	25.06	31.61	26.24	63.72	44.72	73.25	68.00	80.04	64.90	83.16	93.09	64.91	78.54	43.28	73.55	60.94
sd	1.40	3.66	0.77	1.39	3.19	1.67	2.72	1.10	1.84	1.41	0.83	3.02	1.40	2.56	1.72	0.71
$\Delta$	$-21.91$	$-20.90$	$-12.86$	$-9.49$	$-9.77$	$-4.82$	$-8.05$	$+1.77$	$-7.60$	$-0.33$	$+1.57$	$-11.04$	$+0.71$	$-18.12$	$+0.52$	$-8.02$
C100	gaus	shot	impu	defo	glas	moti	zoom	snow	fros	fog	brit	cont	elas	pixe	jpeg	Average

Appendix

Table 18: Per-corruption accuracy without

L_{\mathrm{align}}

, episodic, class-blocked, ten seeds. The line below each result is the standard deviation over seeds.

Figure 2: Accuracy against batch size on IN-RN50, episodic, averaged over ten seeds.

shuffled	gaus	shot	impu	defo	glas	moti	zoom	snow	fros	fog	brit	cont	elas	pixe	jpeg	Average
$\lambda=K/100$	56.14	60.55	41.11	80.25	59.54	82.02	82.34	80.75	76.99	84.44	91.32	80.58	77.84	73.40	73.43	73.38
sd	0.30	0.34	0.26	0.40	0.40	0.31	0.19	0.19	0.31	0.24	0.09	0.19	0.08	0.55	0.16	0.08
$\lambda=10$	43.82	48.71	29.60	67.75	50.51	71.22	70.63	72.90	68.64	76.01	86.84	67.87	68.48	58.96	63.57	63.03
sd	0.28	0.60	0.59	0.58	0.38	0.53	0.58	0.33	0.34	0.29	0.28	0.52	0.40	0.75	0.28	0.21
$\Delta$	$-12.32$	$-11.84$	$-11.51$	$-12.50$	$-9.03$	$-10.80$	$-11.71$	$-7.85$	$-8.35$	$-8.43$	$-4.48$	$-12.71$	$-9.36$	$-14.44$	$-9.86$	$-10.35$
class-blocked	gaus	shot	impu	defo	glas	moti	zoom	snow	fros	fog	brit	cont	elas	pixe	jpeg	Average

Appendix

Table 19: C10, episodic, ten seeds, with

\lambda

set by the rule (

K/100=0.1

) or fixed at the ImageNet value

10

. The line below each result is the standard deviation over seeds.

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

When Test-Time Adaptation Helps, Harms, or Becomes Inactive: A Condition-Level Study on CIFAR-10-C
Aug 23, 2026Sreeja Guha Majumdar, Aratrika SahaTest-Time Adaptation Corruption Robustness
Heritage Institute of Technology Kolkata, West Bengal, India
SPeaR: Test-Time Adaptation with Steering Primitives for Realigning Representations
Sep 21, 2026Muhammad Sudipto Siam Dip, Ali EtemadTest-Time Adaptation Continual Test-Time Adaptation
Department of Electrical and Computer Engineering Queen’s University, Canada
Continual Test-Time Adaptation in Computer Vision: Methods, Benchmarks, and Future Directions
Jul 9, 2026Sarthak Kumar Maharana, Shambhavi Mishra, Yunbei Zhang +6Source-Free Domain Adaptation Distribution Shift Robustness
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

Aug 23, 2026J/K move · Enter open · S save

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

Sreeja Guha Majumdar, Aratrika Saha

Heritage Institute of Technology Kolkata, West Bengal, India

Test-Time Adaptation Corruption Robustness

Open paper AskPDF arXiv

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.

Abstract

Test-Time Adaptation

Explore similar work

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

SPeaR: Test-Time Adaptation with Steering Primitives for Realigning Representations

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