Efficient Test-Time Adaptation (ETTA)

• ETTA is a resource-aware framework that adapts pretrained models during inference under distribution shifts while accounting for latency, memory, and compute constraints.
• It leverages diverse methods such as selective sample updates, lightweight parameter tuning, and training-free mechanisms to balance accuracy and efficiency.
• ETTA techniques show practical gains by reducing backward passes and memory usage, improving effective accuracy in real-time, streaming evaluation settings.

Efficient Test-Time Adaptation (ETTA) denotes the efficiency-oriented branch of test-time adaptation (TTA), in which a pretrained model is adapted to distribution shift using unlabeled test data while explicitly accounting for latency, memory, compute, parameter-update budget, or query cost. In the surveyed literature, ETTA is not a single algorithmic family but a collection of evaluation protocols, optimization strategies, and deployment-aware systems spanning image classification, vision-LLMs, speech foundation models, MRI reconstruction, facial expression recognition, LLMs, reasoning, and black-box APIs. The term “ETTA” is also used as the name of a specific vision-language method based on dynamic embedding updates (Dastmalchi et al., 7 Aug 2025).

1. Definition and problem setting

The core TTA objective is consistent across the literature: adapt a deployed model at inference time under distribution shift without target labels. ETTA narrows this objective by treating adaptation cost as part of the problem definition rather than as an implementation detail. This emphasis appears in several forms: reduction of backward passes, restriction of trainable parameters, avoidance of backpropagation altogether, memory-efficient architectural design, constant-time or near-constant-time online updates, and protocols that penalize slow methods under streaming constraints (Alfarra et al., 2023).

A central theme is that conventional TTA metrics can be misleading when they ignore deployment constraints. The online protocol of “Evaluation of Test-Time Adaptation Under Computational Time Constraints” formalizes this point by modeling a constant-speed stream and showing that adaptation speed changes method rankings (Alfarra et al., 2023). Subsequent work extends the same efficiency criterion into other bottlenecks: backward-pass count in EATA (Niu et al., 2022), activation memory in EcoTTA (Song et al., 2023), query complexity in BETA (Zhang et al., 17 Apr 2026), GPU memory in E-BATS (Dong et al., 8 Jun 2025), and edge-compatible forward-only adaptation in TED (Luo et al., 13 Oct 2025).

The literature also uses ETTA in a broader architectural sense. Some methods still perform gradient-based updates, but only on a narrow subset of parameters or only on selected samples. Others are training-free or backpropagation-free and instead rely on caches, prototypes, low-dimensional latent coordinate updates, prompt vectors, or context-to-weight mappings. A plausible implication is that ETTA is best understood as a design regime defined by resource-aware adaptation rather than by any single loss or parameterization.

2. Time-aware evaluation and the speed–accuracy trade-off

The most explicit efficiency formalization is the online evaluation protocol proposed for TTA under computational time constraints (Alfarra et al., 2023). In that protocol, data arrives from a stream at constant speed $r$, and each TTA method $g$ has a relative adaptation complexity

$$\mathcal{C}(g)=\left\lceil \frac{r}{R(g)} \right\rceil,$$

where $R(g)$ is the speed of method $g$. If a method is $k$ times slower than the stream, it can adapt only every $k^{\text{th}}$ sample; intermediate samples are processed without adaptation. This differs from the standard offline protocol, which effectively lets the data wait for the model (Alfarra et al., 2023).

Empirical estimates in that study make the efficiency gap concrete. The reported average $\mathcal{C}(g)$ values are $1$ for AdaBN, $3$ for TENT, $g$0 for TTAC-NQ, $g$1 for MEMO, and $g$2 for DDA (Alfarra et al., 2023). Under this online protocol, ranking reversals occur: SHOT from 2020 outperforms SAR from 2023, and simple or fast approaches can exceed the accuracy of slower methods once adaptation speed is priced into the evaluation (Alfarra et al., 2023). One reported example is especially direct: offline, SAR achieves $g$3 error and SHOT $g$4; online, SAR rises to $g$5 while SHOT remains at $g$6 (Alfarra et al., 2023).

This result addresses a common misconception that ETTA is merely “faster TTA” with lower accuracy. The online benchmark instead shows that under realistic streaming assumptions, efficiency can improve effective accuracy because a fast method adapts on more of the stream. The same paper therefore recasts ETTA as an evaluation problem as much as a method-design problem: practical TTA must be both accurate and efficient (Alfarra et al., 2023).

3. Core algorithmic patterns for efficiency

A large fraction of ETTA methods achieve efficiency by reducing when, where, or how adaptation occurs. The dominant patterns are selective updating, lightweight adaptation modules, and adaptation routing based on batch or sample properties.

Efficiency lever	Representative methods	Reported mechanism
Selective sample updates	EATA, DATTA	Reliability/diversity filtering; fine-tune only selected batches or samples
Lightweight parameter updates	EcoTTA, FIESTA	Meta networks or Fisher-selected parameters
Stability-oriented regularization	EcoTTA, LATTA	Self-distilled regularization; EMA anchoring
Backprop reduction	EATA	Skip unreliable or redundant samples
Diversity-aware routing	DATTA	Batch diversity discrimination selects normalization and fine-tuning strategy

EATA is one of the canonical ETTA formulations. It observes that not all test samples contribute equally and proposes active sample selection to identify reliable and non-redundant samples, together with a Fisher regularizer to protect important parameters (Niu et al., 2022). Only selected samples participate in entropy-minimization updates, and only batch-norm affine parameters are adapted. The objective is

$g$7

where $g$8 is the sample score and $g$9 is the Fisher regularizer (Niu et al., 2022). On ImageNet-C level 5, EATA reports about $$\mathcal{C}(g)=\left\lceil \frac{r}{R(g)} \right\rceil,$$0 backward passes versus $$\mathcal{C}(g)=\left\lceil \frac{r}{R(g)} \right\rceil,$$1 for Tent and more than $$\mathcal{C}(g)=\left\lceil \frac{r}{R(g)} \right\rceil,$$2 million for MEMO, while achieving about $$\mathcal{C}(g)=\left\lceil \frac{r}{R(g)} \right\rceil,$$3 error versus $$\mathcal{C}(g)=\left\lceil \frac{r}{R(g)} \right\rceil,$$4 for Tent and $$\mathcal{C}(g)=\left\lceil \frac{r}{R(g)} \right\rceil,$$5 for MEMO (Niu et al., 2022).

EcoTTA shifts the efficiency target from backward-pass count to memory footprint. Its backbone remains frozen and lightweight meta networks are attached to encoder partitions, so only the meta networks require activation storage for backpropagation. A self-distilled regularization term constrains the meta-network outputs not to deviate significantly from the frozen source outputs, thereby addressing catastrophic forgetting and error accumulation without additional memory (Song et al., 2023). Reported memory reductions relative to CoTTA reach $$\mathcal{C}(g)=\left\lceil \frac{r}{R(g)} \right\rceil,$$6 for ResNet-50 and $$\mathcal{C}(g)=\left\lceil \frac{r}{R(g)} \right\rceil,$$7 for WideResNet-40 while retaining or improving performance (Song et al., 2023).

DATTA introduces a different routing principle: batch diversity. Its Diversity Discrimination, Diversity Adaptive Batch Normalization, and Diversity Adaptive Fine-Tuning modules distinguish high-diversity from low-diversity batches and then select normalization and fine-tuning strategies accordingly (Ye et al., 2024). This is explicitly efficiency-motivated: fine-tuning is skipped when it is likely to be harmful or wasteful. The paper reports up to a $$\mathcal{C}(g)=\left\lceil \frac{r}{R(g)} \right\rceil,$$8 increase in accuracy compared to state-of-the-art methodologies and cites per-batch latency of $$\mathcal{C}(g)=\left\lceil \frac{r}{R(g)} \right\rceil,$$9s for DATTA versus $R(g)$0s for NOTE on CIFAR10-C (Ye et al., 2024).

LATTA treats instability itself as an efficiency problem. It augments entropy-based TTA with SGLD-inspired noisy weight perturbation and an EMA weight anchor, without architectural changes or Monte Carlo passes (Vejendla, 7 Oct 2025). The reported CIFAR-10-C accuracy is $R(g)$1, compared with $R(g)$2 for EATA, and the method is described as simultaneously improving average accuracy and reducing performance variance (Vejendla, 7 Oct 2025).

4. Training-free, cache-based, and forward-only ETTA

A major ETTA trajectory avoids parameter-gradient adaptation altogether. In these methods, efficiency comes from training-free inference-time structures such as caches or prototypes, or from forward-only optimization in a compact subspace.

TDA is an early reference point for training-free adaptation in vision-LLMs. It maintains a lightweight key-value cache, progressively refines pseudo labels, and adds negative pseudo labeling to address pseudo-label noise, all without backpropagation (Karmanov et al., 2024). On ImageNet validation with $R(g)$3 images and one GPU, TDA is reported at $R(g)$4 minutes, compared with $R(g)$5h $R(g)$6min for TPT and $R(g)$7h $R(g)$8min for DiffTPT, while the base CLIP model runs in $R(g)$9 minutes (Karmanov et al., 2024).

Later vision-language work compresses or replaces the cache. The method named ETTA introduces a Recursive Updating module that integrates all incoming test samples into contextual embeddings without storing an unbounded cache, and an Adaptive Ensemble module that dynamically filters prompts per class (Dastmalchi et al., 7 Aug 2025). For $g$0 ImageNet images, it is reported at $g$1 minutes, compared with $g$2 minutes for TDA, $g$3 hours for TPT, and $g$4 hours for DiffTPT (Dastmalchi et al., 7 Aug 2025). PTA replaces cache retrieval with class-specific knowledge prototypes updated by adaptive EMA. On ImageNet-1K it reports $g$5 FPS, or $g$6 of CLIP’s inference speed, with $g$7 MiB memory; the same comparison reports TDA at $g$8 FPS ($g$9 of CLIP), ADAPT at $k$0 FPS, and TPT at $k$1 FPS (Huang et al., 23 Apr 2026). On $k$2 cross-domain benchmarks, PTA improves CLIP from $k$3 to $k$4, whereas TDA reaches $k$5 (Huang et al., 23 Apr 2026).

For Vision Transformers, NAVIA addresses the interaction between token aggregation and TTA. The paper formalizes ETTA as preserving TTA capability while reducing inference latency, argues from a mutual-information perspective that token aggregation causes information loss that norm-tuning cannot fully recover, and proposes direct augmentation of the [CLS] token embedding together with shallow-layer [CLS] biases (Xiong et al., 5 Aug 2025). Reported experiments show more than a $k$6 inference-latency reduction and over $k$7 improvement over state-of-the-art methods (Xiong et al., 5 Aug 2025).

TED extends forward-only ETTA to edge deployment. It keeps model parameters frozen and performs CMA-ES optimization only over a compact low-dimensional coordinate vector in the latent principal subspace (Luo et al., 13 Oct 2025). The adapted latent is

$k$8

with optimization of $k$9 by entropy minimization (Luo et al., 13 Oct 2025). The paper reports up to $k^{\text{th}}$0 times lower computational complexity, state-of-the-art performance on ImageNet and Google Speech Commands series datasets, and deployment on the ZYNQ-7020 platform (Luo et al., 13 Oct 2025).

5. Modality-specific and deployment-oriented extensions

ETTA has expanded beyond standard image classification, and the efficiency criterion often changes with modality. In speech foundation models, E-BATS argues that vision-derived BP-free methods transfer poorly because speech uses LayerNorm, CNN-plus-transformer encoders, variable-length sequence-to-sequence outputs, batch size $k^{\text{th}}$1, and a dominant blank token under CTC (Dong et al., 8 Jun 2025). Its solution is forward-pass-only prompt adaptation optimized by CMA-ES, a multi-scale loss combining entropy minimization with utterance-level and token-wise alignment, and a test-time EMA across utterances (Dong et al., 8 Jun 2025). Across four noisy speech datasets spanning sixteen acoustic conditions, E-BATS reports $k^{\text{th}}$2 gains over BP-free baselines and $k^{\text{th}}$3–$k^{\text{th}}$4 times GPU memory savings compared with BP-based methods (Dong et al., 8 Jun 2025).

MRI reconstruction imposes different constraints: patient-wise variation, scarce data, and slice-level structure preservation. D2SA addresses these with a dual-stage design: an MR-INR branch for patient-wise distribution adaptation and a single-slice refinement stage using a frozen backbone plus a learnable anisotropic diffusion module (Zhang et al., 25 Mar 2025). On the anatomy shift reported in the paper, DIP-TTT obtains $k^{\text{th}}$5 min, whereas FINE+MR-INR+SST (D2SA) reports $k^{\text{th}}$6 min (Zhang et al., 25 Mar 2025).

Video-based facial expression recognition motivates parameter selection rather than sample selection. FIESTA computes Fisher scores from pseudo-labeled frames and updates only the most important parameters, optionally only $k^{\text{th}}$7 of all parameters or $k^{\text{th}}$8 of early-layer weights (Honarmand et al., 29 Mar 2025). On AffWild2, the reported best variant achieves a $k^{\text{th}}$9 improvement in F1 score over the base model while adapting only $\mathcal{C}(g)$0 parameters, more than $\mathcal{C}(g)$1 times fewer than comparable methods (Honarmand et al., 29 Mar 2025).

LLMs and reasoning systems further widen the ETTA scope. StreamAdapter converts context into parameter updates through context mapping and weight absorption, reducing reliance on in-context demonstrations and yielding constant time complexity regardless of demonstration count (Muhtar et al., 2024). DiSCTT adapts reasoning models by estimating instance-level epistemic uncertainty from agreement among sampled trajectories, then routing high-consensus instances to supervised fine-tuning and low-consensus instances to reinforcement learning (Moradi et al., 5 Mar 2026). The reported result is higher accuracy with reduced variance and substantially lower computation and wall-clock time, including a stated $\mathcal{C}(g)$2 reduction compared to TTRL in one comparison (Moradi et al., 5 Mar 2026).

6. Stability, calibration, monitoring, and black-box adaptation

A persistent criticism of efficiency-driven TTA is that faster adaptation can be less reliable. Several ETTA papers therefore move beyond raw accuracy and treat calibration, forgetting, and failure detection as primary objectives.

EATA-C extends EATA by separating reducible model uncertainty from inherent data uncertainty (Tan et al., 2024). It measures model uncertainty via divergence between the full network and sub-networks, replaces pure confidence pushing with a divergence loss, and applies a min-max entropy regularizer that selectively increases or decreases confidence depending on prediction disagreement (Tan et al., 2024). The paper reports that on ImageNet-C with ResNet-50, EATA-C reduces ECE from $\mathcal{C}(g)$3 for EATA to $\mathcal{C}(g)$4, and on ImageNet-R with ViTBase it improves accuracy by $\mathcal{C}(g)$5 over EATA while attaining the lowest ECE of $\mathcal{C}(g)$6 (Tan et al., 2024).

AETTA addresses a different reliability problem: detecting when TTA is failing without labels. It estimates accuracy from prediction disagreement between the adapted model and dropout inferences, then corrects this estimate under adaptation failures using entropy-based scaling (Lee et al., 2024). Across four baselines and six TTA methods, AETTA is reported to yield an average of $\mathcal{C}(g)$7p more accurate estimation than the baselines, and a model-recovery case study shows an $\mathcal{C}(g)$8 percentage-point improvement in average classification accuracy (Lee et al., 2024). This suggests that ETTA increasingly includes self-monitoring mechanisms, not just efficient updates.

Energy-based methods address overconfidence from another angle. EpoTTA replaces sampling-heavy energy-based adaptation with a sampling-free preference objective derived from a residual-energy parameterization mathematically equivalent to DPO (Han et al., 26 May 2025). On TinyImageNet-C, the paper reports $\mathcal{C}(g)$9K GFLOPs for EpoTTA versus $1$0K GFLOPs for TEA, while on severity $1$1 TinyImageNet-C it reports $1$2 accuracy for EpoTTA, compared with $1$3 for TEA, $1$4 for TENT, and $1$5 for SHOT (Han et al., 26 May 2025). The stated motivation is calibration without the impracticality of SGLD sampling (Han et al., 26 May 2025).

Some work relaxes the standard unsupervised assumption. Active Test-Time Adaptation introduces ATTA, in which a limited budget of labeled test instances is queried online and combined with entropy-balanced pseudo-source samples to avoid catastrophic forgetting (Gui et al., 2024). Its SimATTA algorithm is described as efficient and is reported at $1$6s per $1$7 update steps on Tiny-ImageNet-C, versus $1$8s–$1$9s for TTA methods, while achieving large gains on PACS and VLCS with a budget of $3$0 active labels (Gui et al., 2024). This does not fit the fully unlabeled TTA setting, but it is part of the broader ETTA discussion because it treats annotation budget and computational overhead jointly.

Finally, BETA extends ETTA to strict black-box models accessible only through APIs. It uses a lightweight local steering model to create a gradient pathway, harmonizes local and remote predictions, applies consistency regularization and prompt-learning-oriented filtering, and requires only one API call per test sample (Zhang et al., 17 Apr 2026). On ImageNet-C, the reported gains are $3$1 accuracy on ViT-B/16 and $3$2 on CLIP, and on a commercial API the method is described as comparable to ZOO at $3$3x lower cost while maintaining real-time inference speed (Zhang et al., 17 Apr 2026).

Taken together, these studies define ETTA as a deployment-centered reformulation of TTA. The shared conclusion is not that adaptation should be weakened for speed, but that efficiency constraints change what counts as a strong adaptation method. In realistic settings, the decisive variables include stream rate, backward-pass count, memory footprint, prompt or cache size, active-label budget, API-query budget, calibration quality, and failure detectability as much as benchmark accuracy itself (Alfarra et al., 2023).