SATOR: Backpropagation-Free 3D Test-Time Adaptation via Sample-Adaptive Token Routing

Inseok Kong; Geunyoung Jung; Jiyoung Jung

cs.CVOct 15, 2026

SATOR: Backpropagation-Free 3D Test-Time Adaptation via Sample-Adaptive Token Routing

Authors: Inseok Kong, Geunyoung Jung, Jiyoung Jung

Organizations: Department of Geo Informatics, University of Seoul, South Korea · Department of Artificial Intelligence, University of Seoul, South Korea

Abstract

Test-time adaptation (TTA) mitigates the performance degradation of source-pretrained models on unlabeled test data under distribution shifts. Recent backpropagation-free token-purging methods address 3D point cloud distribution shifts by filtering unreliable tokens. However, their reliance on predefined purge-size candidates requires multiple forward passes per sample and increases inference cost. We propose SATOR (Sample-Adaptive TOken Routing), a source-free and backpropagation-free 3D TTA framework that eliminates this search by deriving both the partitioning threshold and routing decision directly from each sample. SATOR computes token abnormality scores from complementary attention and geometric relation graphs. It then derives a sample-specific partition into retained and pruned tokens via Otsu-based separability, without evaluating candidate sizes. Keep-ratio-calibrated routing then selects the pruned prediction only when its entropy, penalized by the keep ratio, is lower than that of the original prediction. Experiments on ModelNet40-C, ScanObjectNN-C, and ShapeNet-C demonstrate that SATOR achieves state-of-the-art accuracy on three 3D corruption benchmarks.

Figures & tables

Figure 1: Overall pipeline of SATOR . Token abnormality scores are computed from attention and geometric relation graphs and adaptively fused based on their Otsu separability. A sample-specific token partition is then derived directly from the fused score distribution via sorted-split Otsu thresholding, eliminating iterative purge-size search. A keep-ratio-calibrated routing mechanism selects between the original and pruned token sequence predictions, accounting for the keep ratio in the routing criterion.

Test-Time Training
Method	Venue	UNI	GAUSS	BACKG	IMPUL	UPSAM	RBF	RBF-INV	DEN-DEC	DENS-INC	SHEAR	ROT	CUT	DISTORT	OCCL	LIDAR	MEAN	DIFF
SOURCE	-	66.53	59.12	7.21	31.69	74.72	67.71	69.77	62.28	75.08	74.35	38.13	58.63	70.02	38.53	18.64	54.16	0.00
MATE [ 12 ]	ICCV’23	69.80	61.80	18.90	63.90	72.50	64.00	66.00	74.00	80.80	71.00	36.70	69.20	66.30	38.40	29.90	58.88	4.72
SMART-PC [ 1 ]	ICML’25	82.40	80.10	12.00	67.10	84.50	76.00	78.60	67.30	72.90	73.30	43.90	72.60	73.50	37.40	24.80	63.09	8.93
Backpropagation Method
PL [ 8 ]	ICML-W’13	73.74	63.70	10.53	55.23	62.56	67.18	68.35	74.15	78.28	69.69	52.35	72.45	65.11	43.40	37.12	59.59	5.43

Table 1: Top-1 classification accuracy (%) across all 15 corruption types in ModelNet40-C. The highest mean accuracy is highlighted in bold and the second highest is underlined .

Test-Time Training
Method	Venue	UNI	GAUSS	BACKG	IMPUL	UPSAM	RBF	RBF-INV	DEN-DEC	DENS-INC	SHEAR	ROT	CUT	DISTORT	OCCL	LIDAR	MEAN	DIFF
SOURCE	-	77.34	71.78	8.64	54.51	77.90	75.50	76.06	85.25	76.42	80.48	57.08	85.12	76.05	10.97	7.13	61.35	0.00
MATE [ 12 ]	ICCV’23	77.80	74.70	4.30	66.20	78.60	76.30	75.30	86.10	86.60	79.20	56.10	84.10	76.10	12.30	13.10	63.12	1.77
SMART-PC [ 1 ]	ICML’25	80.80	78.90	8.90	60.40	81.80	81.10	81.70	84.80	78.40	80.80	63.70	84.90	79.80	11.50	8.80	64.42	3.07
Backpropagation Method
PL [ 8 ]	ICML-W’13	69.70	66.35	8.24	59.37	62.91	71.11	72.40	80.50	79.18	74.18	56.92	80.43	71.05	9.07	9.98	58.09	-3.26

Table 2: Top-1 classification accuracy (%) across all 15 corruption types in ShapeNet-C. The highest mean accuracy is highlighted in bold and the second highest is underlined .

Test-Time Training
Method	Venue	UNI	GAUSS	BACKG	IMPUL	UPSAM	RBF	RBF-INV	DEN-DEC	DENS-INC	SHEAR	ROT	CUT	DISTORT	OCCL	LIDAR	MEAN	DIFF
SOURCE	-	21.69	18.76	16.70	18.42	22.20	45.96	46.99	71.95	69.36	48.88	35.63	72.98	49.40	6.71	8.78	36.96	0.00
MATE [ 12 ]	ICCV’23	27.50	29.40	14.30	22.20	25.60	40.80	42.00	73.70	63.20	45.10	35.30	73.30	45.30	7.10	9.30	36.94	-0.02
SMART-PC [ 1 ]	ICML’25	27.50	39.10	19.30	21.50	29.80	44.20	48.90	68.30	60.20	49.40	45.40	70.10	49.10	8.40	12.20	39.56	2.60
Backpropagation Method
PL [ 8 ]	ICML-W’13	45.44	43.37	30.81	53.36	44.75	60.24	63.34	62.31	66.95	56.11	52.32	64.72	59.04	10.67	11.19	48.31	11.35

Table 3: Top-1 classification accuracy (%) across all 15 corruption types in ScanObjectNN-C. The highest mean accuracy is highlighted in bold and the second highest is underlined .

Table 5

Table 6

Test-Time Adaptation

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

See all 318 papers

ZOTTA: Test-Time Adaptation with Gradient-Free Zeroth-Order Optimization
Mar 15, 2026Ronghao Zhang, Shuaicheng Niu, Qi Deng +3Test-Time Adaptation Zeroth-Order Optimization
South China University of Technology, Guangzhou, 510000, China. · Nanyang Technological University, 639798, Singapore. · Xi’an Jiaotong University, Xi’an, 710049, China. +1
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
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

Mar 15, 2026J/K move · Enter open · S save

ZOTTA: Test-Time Adaptation with Gradient-Free Zeroth-Order Optimization

Ronghao Zhang, Shuaicheng Niu, Qi Deng, Yanjie Dong+2

South China University of Technology, Guangzhou, 510000, China. · Nanyang Technological University, 639798, Singapore. · Xi’an Jiaotong University, Xi’an, 710049, China.+1

Test-Time Adaptation Zeroth-Order Optimization Image Classification

Open paper AskPDF arXiv

Test-time adaptation (TTA) aims to improve model robustness under distribution shifts by adapting to unlabeled test data, but most existing methods rely on backpropagation (BP), which is computationally costly and incompatible with non-differentiable models such as quantized models, limiting practical deployment on numerous edge devices. Recent BP-free approaches alleviate overhead but remain either architecture-specific or limited in optimization capacity to handle high-dimensional models. We propose ZOTTA, a fully BP-free TTA framework that performs efficient adaptation using only forward passes via Zeroth-Order Optimization (ZOO). While ZOO is theoretically appealing, naive application leads to slow convergence under high-dimensional parameter spaces and unstable optimization due to the lack of labels. ZOTTA overcomes these challenges through 1) Distribution-Robust Layer Selection, which automatically identifies and freezes layers that already extract distribution-invariant features, updating only domain-sensitive layers to reduce the optimization dimensionality and accelerate convergence; 2) Spatial Feature Aggregation Alignment, which stabilizes ZOO by aligning globally aggregated spatial features between source and target to reduce gradient variance. Together, these components enable architecture-agnostic and stable BP-free adaptation. Extensive experiments on ImageNet-C/R/Sketch/A show that ZOTTA outperforms or matches BP-based methods, e.g., it reduces memory usage by 84% and improves accuracy by 3.9% over SAR on ImageNet-C. Code:https://github.com/Zhang-Ronghao/zotta.

Abstract

Test-Time Adaptation

Explore similar work

ZOTTA: Test-Time Adaptation with Gradient-Free Zeroth-Order Optimization

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

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