Hidden Trajectories Don't Lie: Correcting Corruption Drift for Test-Time Adaptation

Amirhossein Sabour; Amir Khasahmadi; Mengliu Zhao

cs.LGOct 16, 2026

Hidden Trajectories Don't Lie: Correcting Corruption Drift for Test-Time Adaptation

Authors: Amirhossein Sabour, Amir Khasahmadi, Mengliu Zhao

Organizations: Autodesk Research · McMaster University

Abstract

Test-time adaptation (TTA) adapts pretrained models to distribution shifts encountered during deployment. However, most existing methods rely on the weight representations of the final layer or signals derived from the model's output probability distribution. We study hidden trajectories: the ordered sequence of intermediate representations produced across a deep neural network. We observe that clean and corrupted trajectories become more separable as corruption severity increases. We introduce Drift-TTA, which adds zero-initialized low-rank adapters to a frozen backbone and trains them to align corrupted hidden trajectories with those from cleaner inputs. The approach supports two variants: a source-dependent version trained with clean targets, and a source-free setting that learns from corrupted images at different severities. We develop an online TTA approach that freezes the backbone and adapters while updating a scalar gate to control adaptation strength. We further cast trajectory alignment as MAP inference under a Gaussian observation model and derive a KL-based transfer bound characterizing test-time trajectory misalignment in terms of offline residual error, residual dimensionality, and distribution shift between offline and test corruptions. We evaluate Drift-TTA on CIFAR-10-C, CIFAR-100-C, and ImageNet-C using supervised ViTs and CLIP models. On CIFAR-100-C at severity 5, our source-free online approach improves accuracy over no TTA by 13.3 percentage points with ViT-B/16 and 11.6 percentage points with CLIP ViT-B/16, reaching 74.9% and 47.4%, respectively. Under the same setting, prior TTA baselines yield ECEs of 18.00% for ViT and 14.70% for CLIP, while our approach reduces them to 9.53% and 5.77%, respectively. These results demonstrate that hidden trajectory alignment provides an effective signal for both offline correction and lightweight online test-time adaptation.

Figures & tables

Figure 1: Corruption progressively shifts hidden trajectories, while Drift-TTA moves them back toward clean. (a) Schematic of severity-dependent trajectory drift and the correction learned by Drift-TTA. (b) Distributions of the trajectory magnitude and angle features for clean and corrupted inputs at severities 1–5, before and after applying Drift-TTA. Each column compares clean inputs with one corruption severity. Without adaptation, the corrupted distributions move farther from clean as severity increases. With Drift-TTA, they shift back toward the clean distribution.

			Test-time setting
Method	Adaptation signal	Modules updated	Access	Granularity	Persistence	Multi-layer
Tent ( Wang et al., 2021 )	Entropy	Norm. affine	Online	Batch	Continual
SAR ( Niu et al., 2023 )	Entropy + sharpness	Norm. affine	Online	Batch	Continual
EATA ( Niu et al., 2022 )	Entropy + Fisher	Norm. affine	Online	Batch	Continual
FOA ( Niu et al., 2024 )	Entropy + feature statistics	Prompt + activations	Online	Batch	Continual	✓
CMF ( Lee and Chang, 2024 )	Entropy + filtering	Model parameters	Online	Batch	Continual

Table 1: Comparison of representative TTA methods. “Adaptation signal” denotes the information used to guide adaptation, while “Modules updated” denotes what is changed during adaptation. We describe the test-time setting along three axes: “Access” indicates whether adaptation uses test samples as they arrive (online) or after they are available together (offline). “Granularity” indicates whether the method can adapt using a single test sample (instance), a group of test samples (batch), or both. “Persistence” indicates whether test-time changes are reset after adaptation (episodic) or kept for later test samples (continual). “Multi-layer” indicates whether representations from multiple network layers are used during adaptation.

	Data required			When adaptation occurs
Paradigm	Clean	Clean	Unlabeled	Before	At test
	images	labels	corrupted images	testing	time
Fine tuning	✓	✓		✓
Unsupervised DA	✓	✓	✓	✓
Test time training	✓	✓	✓		✓
Online TTA (e.g. Tent )			✓		✓

Table 2: Comparison of adaptation paradigms. The table compares the data required beyond the pretrained model and when adaptation is performed. For Drift-TTA, source dependent and source free denote two training settings: source dependent uses clean images as trajectory targets for corrupted inputs, while source free uses corrupted and further corrupted images without clean images or labels. After training, either setting can be used directly with the learned adapters fixed or with Drift-TTA-online, which additionally updates only scalar gates at test time.

Figure 2: Architecture of Drift-TTA. A frozen pretrained backbone is augmented with zero-initialized low-rank adapters attached to selected linear maps. The mean-pooled representations $\mathbf{h}_{1},\ldots,\mathbf{h}_{L}$ produced across the backbone form the hidden trajectory. The learned adapters move corrupted trajectories toward their clean counterparts while keeping the pretrained backbone frozen. After training, the adapters can be used directly with fixed weights. For online adaptation, Drift-TTA-online keeps the backbone and adapter weights frozen and updates only the strength of each learned correction.

		Method	Gaussian	Shot	Impulse	Defocus	Glass	Motion	Zoom	Snow	Frost	Fog	Brightness	Contrast	Elastic	Pixelate	JPEG	Mean
CIFAR-100-C	ViT	ViT-B/16	42.0	42.4	54.1	73.1	47.5	72.8	75.0	80.5	79.7	57.1	86.0	44.0	64.5	47.7	57.5	61.6
		TENT	42.6	43.4	54.8	73.4	47.2	73.0	75.2	80.3	79.5	54.3	86.0	39.8	64.4	46.8	57.2	61.2
		SAR	42.9	44.5	54.0	73.3	47.6	73.0	74.9	79.5	79.1	53.6	85.4	46.3	63.9	46.9	55.4	61.3
		EATA	42.5	43.8	54.5	73.2	47.9	73.0	75.0	80.1	79.7	55.6	85.7	42.0	65.1	48.9	57.2	61.6
		FOA	43.5	49.0	55.3	74.3	55.3	75.8	78.5	81.0	81.9	76.0	87.2	76.2	70.1	55.6	60.7	68.0
		CMF	56.5	65.3	66.0	77.0	61.1	76.4	79.5	81.0	81.7	77.8	86.1	82.9	69.6	68.7	65.7	73.0

Table 3: Test-time adaptation accuracy (%) on CIFAR-100-C at severity 5. Results are reported across all fifteen corruption types for supervised ViT-B/16 and CLIP ViT-B/16. Drift-TTA-online combines a frozen trajectory-repair adapter with online entropy-based gate calibration under source-dependent (SD) and source-free (SF) training. The best result in each column is shown in green and the second best in blue .

	ViT			CLIP
Method / deployment	CIFAR-10-C	CIFAR-100-C	ImageNet-C	CIFAR-10-C	CIFAR-100-C	ImageNet-C
No-TTA	13.51	10.10	11.54	4.87	6.30	4.12
PEA (ViT) / BITTA (CLIP)	11.44	18.00	33.44	13.58	14.70	18.04
Adapter only (SD)	11.72	8.47	17.82	3.33	10.33	4.51
$+$ Online gate (SD)	8.21	8.06	13.03	3.53	5.01	12.89
$\Delta_{\mathrm{SD}}$	$\downarrow$ 3.23	$\downarrow$ 9.94	$\downarrow$ 20.41	$\downarrow$ 10.25	$\downarrow$ 9.69	$\downarrow$ 13.53

Table 4: Mean expected calibration error (ECE, %,

\downarrow

) on the corruption benchmarks at severity 5. PEA and BITTA are the ViT and CLIP baselines, respectively. “Adapter only” uses the frozen Drift-TTA adapter; “+ Online” adds entropy-based gate calibration. For each setting,

\Delta_{\mathrm{SD}}

and

\Delta_{\mathrm{SF}}

select the lower ECE of these two variants and report its reduction from the corresponding baseline in percentage. Green

\downarrow

values indicate lower ECE. Teal, orange, purple, and gray denote no adaptation, source dependent (SD), source free (SF), and the prior TTA baselines.

Figure 3: UMAP of CLIP ViT-B/16 image embeddings on Frost-corrupted CIFAR-10-C at severity 5. Columns show no adaptation, source-free Drift-TTA without and with online TTA, and source-dependent Drift-TTA without and with online TTA. Points are colored by their ground-truth class. Each Drift-TTA adapter is trained without Frost under the leave-one-family-out protocol, and each panel is projected independently using UMAP with Euclidean distance.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

		Method	Gaussian	Shot	Impulse	Defocus	Glass	Motion	Zoom	Snow	Frost	Fog	Brightness	Contrast	Elastic	Pixelate	JPEG	Mean
CIFAR-10-C	ViT	ViT-Tiny	43.8	49.8	69.3	82.1	71.0	82.3	84.6	91.8	89.5	87.5	95.5	83.4	83.2	51.8	81.4	76.5
		TENT	43.8	50.1	69.6	81.6	71.2	82.6	84.6	91.7	89.7	88.0	95.2	80.5	83.5	49.0	81.6	76.2
		SAR	45.4	54.7	70.0	81.9	69.8	81.7	83.9	89.0	87.5	83.6	93.3	83.2	81.7	51.5	79.5	75.8
		EATA	43.8	49.8	69.3	82.1	71.0	82.3	84.6	91.8	89.5	87.5	95.5	83.4	83.2	51.7	81.4	76.5
		FOA	62.7	71.6	77.5	88.2	76.4	87.7	91.1	91.6	90.9	89.1	95.2	85.3	85.2	76.5	81.3	83.3
		CMF	66.2	71.6	71.1	87.8	75.6	87.5	90.5	89.6	90.5	88.1	94.8	86.3	83.3	83.3	83.2	83.3

Appendix

Table 5: Per-corruption test-time adaptation accuracy (%) on CIFAR-10-C for unimodal ViT and CLIP-based classification across the fifteen corruption types. The best result in each column is shown in green and the second best in blue .

		Method	Gaussian	Shot	Impulse	Defocus	Glass	Motion	Zoom	Snow	Frost	Fog	Brightness	Contrast	Elastic	Pixelate	JPEG	Mean
ImageNet-C	ViT	ViT-B/16	56.7	56.8	57.5	46.9	35.6	53.1	44.8	62.2	62.5	65.7	77.6	32.6	46.0	66.9	67.6	55.5
		TENT	57.1	58.1	59.2	44.7	43.2	56.6	50.6	62.8	60.5	65.2	78.0	59.7	49.8	68.2	68.6	58.8
		SAR	59.9	62.2	62.8	54.1	54.1	59.1	54.5	63.5	65.6	65.3	78.2	64.4	58.3	69.2	69.7	62.7
		EATA	57.3	59.1	59.9	53.6	49.4	58.2	51.8	63.0	62.9	65.7	77.8	62.0	55.7	65.8	68.7	60.7
		FOA	61.5	63.5	64.3	56.9	55.1	61.0	60.9	68.4	70.9	73.6	80.9	66.0	61.8	73.5	73.6	66.1
		CMF	60.0	61.2	60.9	56.6	56.8	62.4	60.8	69.3	67.9	72.7	78.8	65.2	69.4	73.8	72.0	65.9

Appendix

Table 6: Per-corruption test-time adaptation accuracy (%) on ImageNet-C for unimodal ViT and CLIP-based classification across the fifteen corruption types. The best result in each column is shown in green and the second best in blue .

		Method	Speckle	Gaussian Blur	Spatter	Saturate	Mean
ImageNet-C	ViT	ViT-B/16	64.3	44.9	69.1	73.4	62.9
		PEA	68.4	50.4	73.9	77.9	67.6
		Drift-TTA (SF)	69.4	53.0	69.7	74.3	66.6
		Drift-TTA (SD)	71.1	54.9	69.9	73.8	67.4
		Drift-TTA-online (SF)	69.8	54.5	69.9	74.5	67.2
		Drift-TTA-online (SD)	71.9	56.9	70.5	74.2	68.4

Appendix

Table 7: Test-time adaptation accuracy (%) on ImageNet-C for the four held-out (unseen) corruption families at severity 5. A single Drift-TTA adapter is trained on the fifteen standard corruption families and evaluated, without further training, on four disjoint validation corruptions. “Drift-TTA” is the frozen adapter (offline); “Drift-TTA-online” adds online entropy-based gate calibration; SD and SF denote source-dependent and source-free training. The best result in each column (within each backbone) is shown in green and the second best in blue .

		Method	Speckle	Gaussian Blur	Spatter	Saturate	Mean
CIFAR-10-C	ViT	ViT-Tiny	54.8	77.5	94.7	94.0	80.3
		PEA	70.3	87.0	95.1	94.7	86.8
		Drift-TTA (SF)	76.6	92.8	95.0	94.6	89.8
		Drift-TTA (SD)	88.4	94.1	95.3	94.1	93.0
		Drift-TTA-online (SF)	81.6	93.5	94.6	94.2	91.0
		Drift-TTA-online (SD)	88.3	93.8	95.1	94.1	92.8

Appendix

Table 8: Test-time adaptation accuracy (%) on CIFAR-10-C for the four held-out (unseen) corruption families at severity 5. A single Drift-TTA adapter is trained on the fifteen standard corruption families and evaluated, without further training, on four disjoint validation corruptions. “Drift-TTA” is the frozen adapter (offline); “Drift-TTA-online” adds online entropy-based gate calibration; SD and SF denote source-dependent and source-free training. The best result in each column (within each backbone) is shown in green and the second best in blue .

		Method	Speckle	Gaussian Blur	Spatter	Saturate	Mean
CIFAR-100-C	ViT	ViT-B/16	38.8	70.6	85.3	80.7	68.9
		PEA	64.3	80.9	87.3	85.5	79.5
		Drift-TTA (SF)	71.1	83.9	87.1	84.2	81.6
		Drift-TTA (SD)	80.7	84.4	86.4	81.8	83.4
		Drift-TTA-online (SF)	74.6	84.1	87.1	84.5	82.6
		Drift-TTA-online (SD)	80.3	84.6	86.9	81.9	83.4

Appendix

Table 9: Test-time adaptation accuracy (%) on CIFAR-100-C for the four held-out (unseen) corruption families at severity 5. A single Drift-TTA adapter is trained on the fifteen standard corruption families and evaluated, without further training, on four disjoint validation corruptions. “Drift-TTA” is the frozen adapter (offline); “Drift-TTA-online” adds online entropy-based gate calibration; SD and SF denote source-dependent and source-free training. The best result in each column (within each backbone) is shown in green and the second best in blue .

	Method / deployment	Gaussian	Shot	Impulse	Defocus	Glass	Motion	Zoom	Snow	Frost	Fog	Brightness	Contrast	Elastic	Pixelate	JPEG	Mean
ViT	No-TTA	11.33	8.24	3.45	12.72	9.17	12.84	14.46	15.96	17.11	22.07	14.33	15.92	17.84	13.33	13.86	13.51
	PEA	3.59	3.75	5.54	14.32	7.73	14.02	15.82	13.42	14.81	15.89	13.71	14.36	13.69	10.94	10.08	11.44
	Drift-TTA: adapter only (SD)	10.11	9.97	4.92	13.85	9.55	12.50	14.64	12.68	14.71	16.82	12.35	13.61	13.69	6.17	10.30	11.72
	Drift-TTA: adapter $+$ online gate (SD)	6.26	6.60	6.16	8.74	6.00	7.92	9.66	8.72	10.68	11.34	8.78	8.89	8.73	8.13	6.51	8.21
	$\Delta$ best SD vs. PEA	$\uparrow$ 2.67	$\uparrow$ 2.85	$\downarrow$ 0.62	$\downarrow$ 5.58	$\downarrow$ 1.73	$\downarrow$ 6.10	$\downarrow$ 6.16	$\downarrow$ 4.70	$\downarrow$ 4.13	$\downarrow$ 4.55	$\downarrow$ 4.93	$\downarrow$ 5.47	$\downarrow$ 4.96	$\downarrow$ 4.77	$\downarrow$ 3.57	$\downarrow$ 3.23
	Drift-TTA: adapter only (SF)	12.02	9.88	8.88	16.73	16.31	14.94	16.49	14.85	18.06	19.98	13.53	17.49	16.73	7.79	13.74	14.50

Appendix

Table 10: Expected calibration error (ECE, %,

\downarrow

) on CIFAR-10-C at severity 5. All 15 corruption families are reported for both backbones. PEA is the ViT prior-TTA baseline and BITTA is the CLIP prior-TTA baseline. Drift-TTA is shown with its adapter alone and with additional online gate tuning, under source-dependent (SD) and source-free (SF) training. In each

\Delta

row, the lower ECE of the two variants is selected independently per column and subtracted from PEA/BITTA. Green

\downarrow

values denote how many percentage points the selected ECE is below the baseline; red

\uparrow

values denote how many points it is above. Method-cell tints follow the visualization palette: teal for no adaptation, orange for SD, purple for SF, and gray for PEA/BITTA.

	Method / deployment	Gaussian	Shot	Impulse	Defocus	Glass	Motion	Zoom	Snow	Frost	Fog	Brightness	Contrast	Elastic	Pixelate	JPEG	Mean
ViT	No-TTA	16.84	18.22	9.46	7.62	19.09	6.03	7.31	5.53	7.23	5.37	6.30	10.77	4.40	16.24	11.11	10.10
	PEA	17.69	17.49	22.72	14.10	18.13	16.91	13.93	16.51	16.74	23.06	13.98	20.14	22.56	19.43	16.64	18.00
	Drift-TTA: adapter only (SD)	6.54	7.49	6.61	8.24	5.23	7.75	8.69	9.86	10.32	12.94	7.99	16.79	7.91	5.93	4.75	8.47
	Drift-TTA: adapter $+$ online gate (SD)	9.08	8.41	8.10	7.84	9.38	7.26	7.81	7.19	7.90	6.37	7.09	5.28	7.50	10.41	11.25	8.06
	$\Delta$ best SD vs. PEA	$\downarrow$ 11.15	$\downarrow$ 10.00	$\downarrow$ 16.11	$\downarrow$ 6.26	$\downarrow$ 12.90	$\downarrow$ 9.65	$\downarrow$ 6.12	$\downarrow$ 9.32	$\downarrow$ 8.84	$\downarrow$ 16.69	$\downarrow$ 6.89	$\downarrow$ 14.86	$\downarrow$ 15.06	$\downarrow$ 13.50	$\downarrow$ 11.89	$\downarrow$ 9.94
	Drift-TTA: adapter only (SF)	6.91	4.98	12.48	8.40	5.92	8.47	9.40	9.55	13.70	18.48	9.56	22.20	7.57	6.70	5.08	9.96

Appendix

Table 11: Expected calibration error (ECE, %,

\downarrow

) on CIFAR-100-C at severity 5. All 15 corruption families are reported for both backbones. PEA is the ViT prior-TTA baseline and BITTA is the CLIP prior-TTA baseline. Drift-TTA is shown with its adapter alone and with additional online gate tuning, under source-dependent (SD) and source-free (SF) training. In each

\Delta

row, the lower ECE of the two variants is selected independently per column and subtracted from PEA/BITTA. Green

\downarrow

values denote how many percentage points the selected ECE is below the baseline; red

\uparrow

values denote how many points it is above. Method-cell tints follow the visualization palette: teal for no adaptation, orange for SD, purple for SF, and gray for PEA/BITTA.

	Method / deployment	Gaussian	Shot	Impulse	Defocus	Glass	Motion	Zoom	Snow	Frost	Fog	Brightness	Contrast	Elastic	Pixelate	JPEG	Mean
ViT	No-TTA	12.85	7.46	11.68	9.99	4.99	2.68	4.68	4.40	20.98	48.11	8.98	20.51	5.97	4.73	5.15	11.54
	PEA	36.07	34.25	34.59	31.67	36.02	33.91	39.18	33.59	38.87	40.01	18.93	37.44	37.93	23.61	25.48	33.44
	Drift-TTA: adapter only (SD)	21.28	20.22	21.28	18.71	9.40	19.38	11.17	15.51	18.80	33.80	14.82	25.63	7.65	16.92	12.75	17.82
	Drift-TTA: adapter $+$ online gate (SD)	16.22	14.85	16.31	13.41	4.16	14.25	5.60	10.99	14.06	29.33	11.86	19.08	2.57	12.98	9.74	13.03
	$\Delta$ best SD vs. PEA	$\downarrow$ 19.85	$\downarrow$ 19.40	$\downarrow$ 18.28	$\downarrow$ 18.26	$\downarrow$ 31.86	$\downarrow$ 19.66	$\downarrow$ 33.58	$\downarrow$ 22.60	$\downarrow$ 24.81	$\downarrow$ 10.68	$\downarrow$ 7.07	$\downarrow$ 18.36	$\downarrow$ 35.36	$\downarrow$ 10.63	$\downarrow$ 15.74	$\downarrow$ 20.41
	Drift-TTA: adapter only (SF)	19.31	15.02	16.28	15.23	6.41	13.72	8.28	15.60	24.63	43.98	15.29	34.32	6.41	13.53	13.40	17.43

Appendix

Table 12: Expected calibration error (ECE, %,

\downarrow

) on ImageNet-C at severity 5. All 15 corruption families are reported for both backbones. PEA is the ViT prior-TTA baseline and BITTA is the CLIP prior-TTA baseline. Drift-TTA is shown with its adapter alone and with additional online gate tuning, under source-dependent (SD) and source-free (SF) training. In each

\Delta

row, the lower ECE of the two variants is selected independently per column and subtracted from PEA/BITTA. Green

\downarrow

values denote how many percentage points the selected ECE is below the baseline; red

\uparrow

values denote how many points it is above. Method-cell tints follow the visualization palette: teal for no adaptation, orange for SD, purple for SF, and gray for PEA/BITTA.

Figure 4: Hidden-trajectory drift on CIFAR-100-C. A frozen ViT-B/16 is evaluated on CIFAR-100-C without adaptation. (a) Mean hidden trajectories for clean inputs and Gaussian noise at severities 1–5, projected onto the first two principal components. (b) Relative clean-to-corrupted drift across network depth, averaged over all 15 corruption families. (c) Relative trajectory drift versus top-1 error across the 75 corruption-family and severity conditions (Pearson

r{=}0.73

; Spearman

\rho_{s}{=}0.78

). (d) Top-1 accuracy across severity for four corruption families. (e) Magnitude and angle feature distributions for clean inputs and Gaussian noise at severities 1–5.

Figure 5: Hidden-trajectory drift on ImageNet-C. A frozen ViT-B/16 is evaluated on the full ImageNet-C validation set without adaptation. Panels (a) and (e) use fog as a representative corruption family, while panels (b) – (d) summarize all fifteen corruption families following Figure 4 . Across the $75$ corruption-family and severity conditions, relative trajectory drift is associated with top-1 error (Pearson $r{=}0.78$ ; Spearman $\rho_{s}{=}0.87$ ). The clean point in (c) is shown as a reference and is excluded from both correlations.

Figure 6: UMAP of CLIP ViT-B/16 image embeddings on CIFAR-10-C (severity 5), colored by class. Rows are six corruption families; columns are, left to right, No Adaptation, source-free Drift-TTA without and with the online gate, and source-dependent Drift-TTA without and with the online gate. Column groups are color-coded (teal: no adaptation; purple: source-free; orange: source-dependent). Points are the $L_{2}$ -normalized image embedding used for zero-shot classification, projected with UMAP (Euclidean metric); each panel is embedded independently. Leave-one-family-out: each column uses only the adapter trained without the family shown in that row.

Figure 7: t-SNE of CLIP ViT-B/16 image embeddings on CIFAR-10-C (severity 5), colored by class. Rows are six corruption families; columns are, from left to right, No Adaptation, source-free Drift-TTA without and with online gate adaptation, and source-dependent Drift-TTA without and with online gate adaptation. Points are the $L_{2}$ -normalized image embeddings used for zero-shot classification, projected independently with t-SNE using Euclidean distance and perplexity 30. Layout, color coding, and the leave-one-family-out protocol match Figure 6 .

Figure 8: UMAP of ViT image features on CIFAR-10-C (severity 5), colored by class. Layout, color coding, and protocol match Figure 6 ; here the points are the pre-classifier pooled embedding of the ViT. As with CLIP, both Drift-TTA priors tighten the class structure and the online gate sharpens it further.

Figure 9: t-SNE of ViT image features on CIFAR-10-C (severity 5), colored by class. Layout, color coding, projection settings, and protocol match Figure 7 ; here, the points are the $L_{2}$ -normalized pre-classifier pooled embeddings of the ViT. The class structure follows the same qualitative progression as the corresponding UMAP visualization in Figure 8 .

Panel of Figure 10 — Figure 10: Trajectory-feature distributions, supervised ViT (source-free). Each panel plots $\mathrm{Mag}(H)$ (horizontal) against $\mathrm{Ang}(H)$ (vertical); green is the clean reference and colored contours are severities 1–5 (columns), while the three rows are No Adaptation, offline Drift-TTA, and online Drift-TTA. Top: glass blur on CIFAR-10-C; bottom: shot noise on CIFAR-100-C. Leave-one-family-out.

Panel of Figure 11 — Figure 11: Trajectory-feature distributions, supervised ViT. Green is the clean reference and colored contours are severities 1–5 (columns); the three rows are No Adaptation, offline Drift-TTA, and online Drift-TTA. Top: snow on ImageNet-C (source-free); bottom: pixelate on CIFAR-10-C (source-dependent). Leave-one-family-out.

Panel of Figure 12 — Figure 12: Trajectory-feature distributions, supervised ViT (source-dependent). Green is the clean reference and colored contours are severities 1–5 (columns); the three rows are No Adaptation, offline Drift-TTA, and online Drift-TTA. Top: impulse noise on CIFAR-100-C; bottom: fog on ImageNet-C. Leave-one-family-out.

Panel of Figure 13 — Figure 13: Trajectory-feature distributions, CLIP (source-free). Green is the clean reference and colored contours are severities 1–5 (columns); the three rows are No Adaptation, offline Drift-TTA, and online Drift-TTA. Top: motion blur on CIFAR-10-C; bottom: elastic transform on CIFAR-100-C. Leave-one-family-out.

Panel of Figure 14 — Figure 14: Trajectory-feature distributions, CLIP. Green is the clean reference and colored contours are severities 1–5 (columns); the three rows are No Adaptation, offline Drift-TTA, and online Drift-TTA. Top: zoom blur on ImageNet-C (source-free); bottom: defocus blur on CIFAR-10-C (source-dependent). Leave-one-family-out.

Panel of Figure 15 — Figure 15: Trajectory-feature distributions, CLIP (source-dependent). Green is the clean reference and colored contours are severities 1–5 (columns); the three rows are No Adaptation, offline Drift-TTA, and online Drift-TTA. Top: Gaussian noise on CIFAR-100-C; bottom: contrast on ImageNet-C. Leave-one-family-out.

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

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

Sep 21, 2026

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

Test-time adaptation (TTA) addresses distribution shift using only unlabeled test data. Existing methods typically adapt pretrained models by updating their parameters, limiting both what is adapted and where adaptation can occur within the network. We instead keep the pretrained network frozen and steer its intermediate representations. We introduce SPeaR (Steering Primitive for Realigning Representations), which inserts lightweight learnable modules at stage boundaries and optimizes them directly from the test stream, requiring neither source data nor supervised warm-up. Each primitive is optimized using a gated objective that reduces uncertainty only when adaptation is beneficial, along with a diversity regularizer to prevent collapse, and a multi-depth anchor to stabilize adaptation. We show that steering early representations is the most effective strategy, and that the same primitive transfers across convolutional and Transformer architectures. Across CIFAR-10-C, CIFAR-100-C, and ImageNet-C, SPeaR consistently matches or outperforms methods that adapt orders of magnitude more parameters, remains robust across a wide range of batch sizes, and preserves source-domain performance during continual adaptation.

Muhammad Sudipto Siam Dip, Ali Etemad

Department of Electrical and Computer Engineering Queen’s University, Canada

Test-Time Adaptation Continual Test-Time Adaptation

Ask PDF

Oct 13, 2025

Efficient Test-Time Adaptation through Latent Subspace Coefficients Search

Real-world deployment often exposes vision models to distribution shifts, making test-time adaptation (TTA) important for reliable prediction. Existing TTA methods can be difficult to deploy on edge devices because they require backpropagation, activation storage, or test-time batches, which increase computation, memory, and state. We propose ELaTTA (Efficient Latent Test-Time Adaptation), a forward-only, gradient-free framework for single-instance TTA. ELaTTA freezes the pretrained model and adapts each test input by searching a low-dimensional coefficient vector in a principal latent subspace computed offline from a source-domain cache. The basis is stored once and source data are discarded before deployment. At test time, CMA-ES performs covariance-adaptive Gaussian search, ranks candidate corrections by prediction entropy, and selects the best candidate within a fixed evaluation budget. On five visual benchmarks spanning corruption, natural, rendition, sketch, and cross-domain shifts, ELaTTA provides strong accuracy under strict and continual single-instance protocols. In the strict single-instance comparison, MEMO requires \emph{64.7

\times

} the computation and \emph{12.4

\times

} the peak memory of ELaTTA. A complementary keyword-spotting experiment shows that the same coefficient-search design transfers beyond vision. We further demonstrate ViT-Base adaptation on a Raspberry Pi 5, confirming the feasibility of deploying ELaTTA on an edge device.

Xinyu Luo, Jie Liu, Kecheng Chen +6

Department of Electrical Engineering, City University of Hong Kong, Hong Kong SAR, China

Test-Time Adaptation Continual Test-Time Adaptation

Ask PDF

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

Efficient Test-Time Adaptation through Latent Subspace Coefficients Search