Test-Time Adaptation (TTA)

• Test-Time Adaptation is the process of adapting a pre-trained model during inference using only unlabeled, shifted data without access to the original training set.
• It employs unsupervised objectives such as entropy minimization, pseudo-labeling, and contrastive divergence to update model parameters on-the-fly.
• Applications span various modalities and settings, with performance and stability influenced by factors like online optimization, batch adaptation, and continual domain shift.

Searching arXiv for recent and foundational papers on test-time adaptation to ground the article and verify the provided records. Test-time adaptation (TTA) denotes the adaptation of a pretrained model during inference using only unlabeled test data from a shifted distribution. In the source-free setting emphasized across recent work, the training data are no longer available, and adaptation must proceed from the deployed model alone, often in an online or streaming regime. Across the literature, TTA spans batch-wise, online, continual, episodic, and modality-specific forms, but its common aim is to mitigate target-domain degradation without supervised target labels (Yu et al., 2023, Brahma et al., 2022).

1. Definition, scope, and settings

A standard formulation starts from a source-trained model $f_{\theta_0}$, trained on labeled source data, and a target stream of unlabeled inputs drawn from a shifted distribution. In the benchmark taxonomy, TTA is organized into Test-Time Domain Adaptation (TTDA), Test-Time Batch Adaptation (TTBA), and Online Test-Time Adaptation (OTTA). TTDA assumes the full unlabeled target domain is visible at once; TTBA adapts per batch and discards the adapted parameters afterward; OTTA updates the model incrementally as batches arrive, and continual test-time adaptation (CTTA) extends this to non-stationary target streams (Yu et al., 2023).

Lifelong or continual TTA makes the setting stricter. Rather than a single stationary target domain, the input distribution changes over time, often without explicit domain-boundary signals. PETAL frames this as online lifelong TTA across multiple target domains and emphasizes model drift, error accumulation, catastrophic forgetting, and unreliable uncertainty estimates as central difficulties (Brahma et al., 2022). BP-TTA sharpens the same point in a practical streaming regime by adding class imbalance and strong temporal correlation: consecutive batches may be dominated by a few classes, and continual domain shift and class imbalance interact rather than appearing separately (Huang et al., 30 Jun 2026).

The evaluation literature also distinguishes between algorithmic TTA and deployable TTA. Under a constant-speed data stream, slower methods adapt on fewer samples, so wall-clock latency becomes part of the problem definition rather than a secondary engineering concern (Alfarra et al., 2023).

2. Objectives and what is adapted

A large part of the literature starts from unsupervised objectives defined on target predictions. Entropy minimization is the canonical example: TENT-style methods minimize prediction entropy on unlabeled target inputs, often by updating only normalization parameters, while pseudo-labeling and information-maximization variants adapt either the classifier, the feature extractor, or both (Yu et al., 2023). This family treats TTA as direct optimization on target predictions, but later work repeatedly notes that noisy pseudo-labels and brittle confidence estimates can destabilize updates (Deng et al., 2024, Yuan et al., 2023).

A probabilistic formulation appears in PETAL. There, the source model is treated as an approximate posterior $q(\theta) \approx p(\theta \mid \mathcal{D})$, and test-time adaptation optimizes a self-training objective regularized by the source posterior: $$\log p(\theta \mid \mathcal{D}, \mathcal{U}) = \log q(\theta) - \frac{\bar{\lambda}}{M} \sum_{m=1}^{M} H(y'_m, y_m \mid \bar{x}_m).$$ This yields a student–teacher mechanism, an EMA teacher, and a source-posterior regularizer within a single framework, rather than as disconnected heuristics (Brahma et al., 2022).

An alternative objective family reinterprets the classifier itself. TEA treats a classifier as an energy-based model over inputs with

$$E_\theta(\mathbf{x}) = -\log \sum_y \exp(f_\theta(\mathbf{x})[y]),$$

and adapts the model by lowering the energy of test samples while raising the energy of model-generated negatives through contrastive divergence and SGLD sampling. In that view, TTA targets the model’s implicit marginal $p_\theta(\mathbf{x})$, not only $p_\theta(y\mid \mathbf{x})$ (Yuan et al., 2023).

Recent work also broadens the notion of what can be adapted. MGG keeps the test-time loss but replaces a hand-designed optimizer by a learned meta-optimizer with a Gradient Memory Layer, turning online TTA into a learning-to-optimize problem over noisy unsupervised gradients (Deng et al., 2024). AcTTA keeps baseline TTA objectives unchanged but shifts the trainable parameter set from normalization affine parameters to learnable activation parameters $(c,\lambda_{\text{neg}},\lambda_{\text{pos}})$, thereby moving beyond the prevailing affine-centric view of adaptation (Kim et al., 27 Mar 2026). LATTA instead regularizes the update rule itself via a noisy weight perturbation inspired by SGLD and an EMA weight anchor, so exploration and stability are both embedded in the parameter update (Vejendla, 7 Oct 2025).

3. Mechanistic families

The field has diversified beyond entropy minimization into several recurrent design patterns.

Mechanism	Characteristic signal	Representative papers
Prototype and neighbor self-training	Support sets, prototypes, $k$-NN pseudo-labels	TAST (Jang et al., 2022), ACCUP (Gong et al., 1 Jan 2025), BP-TTA (Huang et al., 30 Jun 2026)
Retrieval-augmented adaptation	External pool $A$ and retrieval encoder $R$	T$q(\theta) \approx p(\theta \mid \mathcal{D})$0AR (Zancato et al., 2023)
Meta/optimizer learning	Meta-trained self-supervision or learned optimizer dynamics	(Ziakas et al., 11 Jun 2025, Deng et al., 2024)
Energy-based adaptation	Input-space energy and contrastive divergence	TEA (Yuan et al., 2023)
Update-rule regularization	Langevin perturbation, EMA anchoring	LATTA (Vejendla, 7 Oct 2025)
Activation-aware adaptation	Learnable activation centers and slopes	AcTTA (Kim et al., 27 Mar 2026)
Episodic generative adaptation	Token-level consensus pseudolabels with per-query reset	TTAdapt (Kaya et al., 3 Oct 2025)

Prototype-based and neighborhood-based methods are especially prominent. TAST defines a nearest-neighbor pseudo-label distribution from support-set neighbors in embedding space, trains lightweight adaptation modules to match that distribution, and predicts with an ensemble average over multiple modules (Jang et al., 2022). ACCUP combines entropy-filtered prototypes, augmentation ensemble, entropy comparison between classifier and prototype predictions, and an augmented contrastive clustering loss specialized to time series (Gong et al., 1 Jan 2025). BP-TTA couples class-balanced memory sampling with evolving prototypes to address continual shift and class imbalance simultaneously (Huang et al., 30 Jun 2026).

Retrieval has emerged as another axis. T$q(\theta) \approx p(\theta \mid \mathcal{D})$1AR augments test-time contrastive adaptation with an external unlabeled pool $q(\theta) \approx p(\theta \mid \mathcal{D})$2 and a retrieval encoder $q(\theta) \approx p(\theta \mid \mathcal{D})$3, so negatives are not limited to synthetic augmentations of the target batch but can include retrieved real samples from a broader data pool (Zancato et al., 2023). This external-memory view is orthogonal to entropy minimization and prototype methods.

4. Modality-specific instantiations

In generative VLMs, TTAdapt is defined as a per-query episodic procedure. TTAug first aggregates token-level predictive distributions across augmented image–text inputs,

$q(\theta) \approx p(\theta \mid \mathcal{D})$4

and TTAdapt then fine-tunes all model parameters on the resulting consensus pseudolabels for a few steps before resetting to the original checkpoint for the next query. On SmolVLM2-2.2B, this produces large gains on open-ended tasks: COCO rises from 9.1 to 16.9 with TTAug and to 35.9 with TTAdapt, while GQA rises from 0.0 to 5.5 and then 13.5. The same table also shows that adaptation can hurt already strong tasks such as OCRBench or TextVQA, so TTAdapt is explicitly not a universally beneficial add-on (Kaya et al., 3 Oct 2025).

In robotics and visuomotor progress estimation, test-time adaptation is used to refine a goal-conditioned value function over trajectories. The model combines a frozen CLIP encoder, a low-dimensional adaptation module, and a progress head, and updates only the adaptation module online using a learned self-supervised reconstruction loss. Meta-training is used so that the test-time update improves semantic progress estimation rather than temporal shortcut exploitation. The TTT-IM variant reports 0.7822 VOC on in-distribution tk_pnp, 0.7246 on 1m_pnp, and 0.8203 on dt_tk_pnp, substantially above CLIP-FT and GVL baselines in the reported comparisons (Ziakas et al., 11 Jun 2025).

In ASR, the central problem is that high-entropy frames are often semantically important rather than disposable. Confidence-Enhanced Adaptation therefore upweights high-entropy non-silent frames instead of filtering them out, and Short-Term Consistency Regularization exploits local temporal coherence after that confidence-enhancement stage. On LibriSpeech with Gaussian noise, average WER goes from 41.6 for the source model to 35.8 for TENT, 28.7 for SUTA, and 28.3 for the proposed method; on sung speech with Wav2vec2 Base, average WER drops from 62.1 to 53.9 (Liu et al., 2023).

In generic time series, ACCUP uses magnitude warping for augmentation ensemble, uncertainty-aware prototypes built from low-entropy support samples, entropy comparison between classifier and prototype outputs, and augmented contrastive clustering on logits. It reports 88.16 macro-F1 on UCIHAR, 95.60 on MFD, and 62.65 on SSC, and also transfers to PACS with 88.33 accuracy for ResNet18 and 90.03 for ResNet50 (Gong et al., 1 Jan 2025).

5. Evaluation, efficiency, and robustness

Benchmarking studies show that TTA performance is strongly setting-dependent. Across CIFAR-10-C, CIFAR-100-C, ImageNet-C, Office-Home, and DomainNet126, no single method is strong across both synthetic-corruption and natural-shift scenarios. PredBN is unexpectedly strong on corruption datasets, whereas TTDA methods such as SHOT, NRC, and AdaContrast dominate on Office-Home and DomainNet126; among OTTA methods, EATA and SAR are repeatedly strong, and on ViT-B/16, adapting LayerNorm parameters is markedly more stable than broader backbone updates (Yu et al., 2023).

Compute-aware evaluation changes rankings again. When adaptation speed is explicitly modeled by a relative adaptation speed $q(\theta) \approx p(\theta \mid \mathcal{D})$5, slower methods adapt on fewer incoming samples. Under this online protocol, SHOT from 2020 can outperform SAR from 2023, AdaBN and BN remain unaffected because $q(\theta) \approx p(\theta \mid \mathcal{D})$6, and very slow methods such as DDA or MEMO can collapse toward source-level performance in realistic streams (Alfarra et al., 2023).

Several papers attack robustness and efficiency directly. MGTTA learns a meta-optimizer and reaches 71.3 on ImageNet-C, 70.2 on ImageNet-R, 53.3 on ImageNet-Sketch, and 56.7 on ImageNet-A, while requiring 125.5 s on 50,000 ImageNet-C images versus 242.7 s for SAR and 1636.7 s for FOA; performance saturates even with 64–128 unlabeled pre-training samples for the optimizer (Deng et al., 2024). LATTA reports 58.31 on CIFAR-10-C, compared with 56.10 for EATA and 51.22 for Tent, and attributes the gain to the combination of Langevin-style noise and EMA anchoring rather than either component alone (Vejendla, 7 Oct 2025). AcTTA, by moving adaptation into activation functions, remains markedly stronger than normalization-only baselines in small-batch regimes; on ImageNet-C with ViT-B/16 and batch size 4, TENT records 89.92 error while $q(\theta) \approx p(\theta \mid \mathcal{D})$7 records 62.07 (Kim et al., 27 Mar 2026).

Efficiency questions also appear in modality-specific systems. For small VLMs, TTAug with 16 augmentations on SmolVLM2-2.2B increases memory from 4.60 GB to 8.75 GB and latency from 1.43 s to 4.77 s per query, while TTAdapt adds a further short fine-tuning loop on top of that budget (Kaya et al., 3 Oct 2025).

6. Limitations, controversies, and emerging directions

The contemporary literature repeatedly emphasizes that TTA gains are conditional rather than unconditional. In generative VLMs, full parameter TTAdapt can degrade tasks where the base model is already strong and well-calibrated, so simpler TTAug or aggregation-weight optimization may be preferable (Kaya et al., 3 Oct 2025). In continual streaming settings, BP-TTA explicitly notes dependence on pseudo-label confidence, and its class-aware memory bank and prototype maintenance introduce additional overhead (Huang et al., 30 Jun 2026). In time series, ACCUP shows that the choice of augmentation is not interchangeable: magnitude warping is consistently useful, while permutation can damage temporal structure (Gong et al., 1 Jan 2025).

Several approaches also impose nontrivial assumptions. PETAL requires an approximate source posterior via SWAG-diagonal and uses Fisher-based restoration to decide which parameters are irrelevant enough to reset, thereby making calibration and lifelong robustness part of the adaptation design (Brahma et al., 2022). MGG requires a small target-like unlabeled set to pre-train the meta-optimizer (Deng et al., 2024). TEA relies on SGLD sampling and assumes roughly stable $q(\theta) \approx p(\theta \mid \mathcal{D})$8, while noting that SGLD is time-consuming and that mild shifts may slightly degrade accuracy (Yuan et al., 2023). The progress-estimation framework is meta-trained on 2,986 expert demos from BridgeData v2 and explicitly notes its dependence on expert trajectories and its online optimization cost (Ziakas et al., 11 Jun 2025).

Taken together, these results suggest that test-time adaptation is best understood not as a single algorithmic template but as a family of inference-time learning procedures whose success depends on objective choice, adaptation locus, pseudo-label reliability, stream structure, and compute budget. The most consequential recent trend is therefore not merely stronger adaptation, but more selective adaptation: episodic versus continual, activation-aware versus affine-only, optimizer-learning versus loss engineering, and modality-specific restructuring of what constitutes a reliable self-supervised signal.