Online Distributional Adaptation

• Online Distributional Adaptation is a framework that updates models with arriving unlabeled data to handle evolving test-time distributions.
• It employs techniques like proxy loss minimization, dynamic evaluation, and label-shift correction to sustain predictive performance.
• Empirical studies demonstrate enhanced accuracy and robustness across domains such as image classification, video segmentation, and language modeling.

Online distributional adaptation denotes the family of methods that modify a deployed model while data arrive from a test-time distribution that may differ from the training distribution and may continue to evolve during deployment. In the standard Online Test-Time Adaptation (OTTA) formulation, an unlabeled stream of small batches $\mathcal{B}_t$ is processed once, the model is updated on the fly through an unsupervised proxy loss, and no revisiting is allowed (Wang et al., 2023). A more structured instance is online label shift, where $Q_t(x|y)=Q_0(x|y)$ remains fixed while the test-time label distribution $Q_t(y)=q_t$ drifts arbitrarily and true labels are never observed (Wu et al., 2021). In LLMs, the analogous operation is dynamic evaluation, in which the parameter vector becomes a time-varying state updated after predicting each token (Rannen-Triki et al., 2024). A probabilistic formulation makes this explicit by positing a transition prior $p(\theta_t|\theta_{t-1})$ and an unlabeled observation model $p(X_t|\theta_t)$ over evolving parameters (Corrales et al., 24 Jun 2026).

1. Formal settings and performance criteria

The online setting is defined by temporal ordering and restricted feedback. One common formulation assumes a pretrained network $f_{\theta^S}$, a source distribution $p_S(x,y)$, and an unlabeled target stream drawn from one or more target distributions $p_T(x,y)$ that generally differ from $p_S$; adaptation proceeds batch by batch through updates of the form $$\theta^t=\theta^{t-1}-\eta\nabla_\theta L(\mathcal{B}_t;\theta^{t-1})$$ using only a surrogate loss computed on the current batch (Wang et al., 2023). In online streams with distribution shifts, meta-learning formulations further distinguish feature-extractor weights $Q_t(x|y)=Q_0(x|y)$0, a supervised-task head $Q_t(x|y)=Q_0(x|y)$1, and a self-supervised head $Q_t(x|y)=Q_0(x|y)$2, and define per-sample inner-loop updates with no replay buffer (Wu et al., 2023).

Several special cases impose additional structure. Under label shift, the learner receives only unlabeled batches $Q_t(x|y)=Q_0(x|y)$3 while $Q_t(x|y)=Q_0(x|y)$4 is fixed or, in the generalized case, fixed only after an unknown feature map $Q_t(x|y)=Q_0(x|y)$5 (Wu et al., 2024). In online conformal inference, the objective is not point prediction but a set-valued predictor $Q_t(x|y)=Q_0(x|y)$6 whose miscoverage should track a target level under nonstationarity, despite the failure of exchangeability (Jun et al., 6 Nov 2025). In resource-constrained link adaptation, the stream is only partially unlabeled: a model predicts online, then the true outcome arrives after a short delay, and adaptation is triggered only when the absolute error exceeds a predefined threshold $Q_t(x|y)=Q_0(x|y)$7 (Ali et al., 14 Apr 2026).

These settings induce distinct evaluation criteria. Online label-shift work measures average regret against the best fixed classifier in hindsight, with

$Q_t(x|y)=Q_0(x|y)$8

where $Q_t(x|y)=Q_0(x|y)$9 (Wu et al., 2021). OTTA studies typically report classification accuracy or error rate, while video segmentation uses mean intersection-over-union (Wu et al., 2023). Dynamic evaluation emphasizes cumulative log-loss, regret relative to static evaluation, average per-token loss in nats/token, and FLOPs (Rannen-Triki et al., 2024). Conformal methods track average interval width and recalibration delay (Jun et al., 6 Nov 2025). This suggests that “adaptation” is not a single objective but a deployment-dependent trade-off among predictive accuracy, calibration, coverage, compute, and reaction time.

2. Major methodological patterns

A large fraction of the literature can be described through the OTTA taxonomy of optimization-based, data-based, and model-based methods. Optimization-based OTTA includes normalization calibration, mean-teacher or consistency methods, and unsupervised loss designs such as entropy minimization, mutual-information maximization, prototype-based objectives, and Sharpness-Aware Minimization (Wang et al., 2023). Data-based OTTA augments or replays data through predefined or learned perturbations and memory banks. Model-based OTTA modifies the network itself, for example by adding adapter branches, replacing heads with cosine-distance or prototype-based heads, or updating prompt parameters.

Within this broad taxonomy, online label-shift adaptation is structurally distinctive because it exploits the invariance of the class-conditional distributions. The key mechanism is confusion-matrix inversion: if $Q_t(y)=q_t$0, then under label shift $Q_t(y)=q_t$1, and the loss can be written as $Q_t(y)=q_t$2; this turns unlabeled online adaptation into an online learning problem over a re-weighting vector (Wu et al., 2021). OLS-OFU extends that template by interleaving a Revised OLS subroutine with self-supervised feature updates and retraining of the final linear layer on the original labeled source data (Wu et al., 2024).

A second cluster of methods replaces entropy minimization by geometry- or statistics-based alignment objectives. CoMM optimizes the cosine similarity between feature vectors and class weight vectors, using a log-ratio loss that promotes alignment with the predicted class and penalizes alignment with all other classes (Chuah et al., 2024). DA-TTA instead pulls each test-batch feature distribution back toward fixed source BatchNorm statistics through a Distribution Alignment loss defined on per-channel means and variances, with an additional domain shift detection mechanism for continual TTA (Wang et al., 2024). POEM shifts the focus from minimizing entropy to matching the distribution of test-time entropy values to that of the source domain, using a betting martingale for shift detection and a squared matching loss derived from a one-dimensional optimal-transport construction (Bar et al., 2024).

A third pattern emphasizes parameter-efficient or wrapper-based adaptation. In preparation-free heterogeneous cooperative perception, BOLT inserts an adaptive plugin immediately before a frozen fusion module and updates it online through ego-as-teacher distillation, requiring only ego predictions without ground-truth labels (Yang et al., 1 May 2026). The black-box Accuracy-Weighted Ensemble (AWE) wraps any Online Learner with only $Q_t(y)=q_t$3 maintained instances through a multi-resolution instance design and cross-validation-through-time (Baby et al., 9 Apr 2025). In language modeling, dynamic evaluation performs online fine tuning directly on test tokens, while LoRA adapters provide a lower-rank alternative with reduced trainable state (Rannen-Triki et al., 2024). LightTune goes further toward deployment constraints by using forward-only fine-tuning, no backpropagation, and threshold-triggered updates (Ali et al., 14 Apr 2026).

3. Update mechanisms and surrogate objectives

Many methods differ less in overall architecture than in how they construct a usable adaptation signal from unlabeled data. In online label shift, the crucial estimator is

$Q_t(y)=q_t$4

whenever the original model $Q_t(y)=q_t$5 predicts $Q_t(y)=q_t$6; one obtains $Q_t(y)=q_t$7, hence an unbiased estimator of both the expected loss and, under mild differentiability assumptions, its gradient (Wu et al., 2021). The resulting online learner operates on a re-weight vector $Q_t(y)=q_t$8 through an adapted classifier

$Q_t(y)=q_t$9

and can be optimized by Online Gradient Descent or Follow-The-History.

In OTTA proper, the generic mechanism is to compute a proxy loss on each arriving sample or batch and immediately update a restricted parameter subset. Meta-learned stream adaptation performs the online step

$p(\theta_t|\theta_{t-1})$0

typically with a self-supervised rotation-prediction loss and fixed supervised head $p(\theta_t|\theta_{t-1})$1 (Wu et al., 2023). OLS-OFU follows a different sequencing rule: first run the OLS correction with the current model, then update the feature extractor by self-supervised learning on the current unlabeled batch, and finally reinitialize and retrain the linear head on $p(\theta_t|\theta_{t-1})$2 before temperature calibration on $p(\theta_t|\theta_{t-1})$3 (Wu et al., 2024). The placement of the feature update after the OLS step is not incidental; it preserves the independence condition needed for unbiased estimation.

Dynamic evaluation uses the supervised next-token loss available at inference time. If $p(\theta_t|\theta_{t-1})$4 denotes the test token stream, then at step $p(\theta_t|\theta_{t-1})$5 the model incurs

$p(\theta_t|\theta_{t-1})$6

computes $p(\theta_t|\theta_{t-1})$7, and performs $p(\theta_t|\theta_{t-1})$8, with AdamW in practice and optional lower update frequency or LoRA-restricted updates (Rannen-Triki et al., 2024). A probabilistic state-space formulation generalizes this intuition: propagate a prior through $p(\theta_t|\theta_{t-1})$9, update it with an unlabeled Gibbs-posterior surrogate $p(X_t|\theta_t)$0, and approximate the resulting posterior by an Extended-Kalman-style Gaussian update in parameter space (Corrales et al., 24 Jun 2026).

Other objectives are explicitly designed to improve robustness to non-i.i.d. streams. DA-TTA minimizes a Distribution Alignment loss based on the absolute deviation between batch-level test means and variances and precomputed source statistics at BatchNorm layers (Wang et al., 2024). CoMM replaces entropy with a cosine-alignment log-ratio over class weights (Chuah et al., 2024). POEM first transforms test entropy values through a source-domain empirical CDF, then applies an adaptive betting function $p(X_t|\theta_t)$1, and finally updates normalization-layer parameters by minimizing $p(X_t|\theta_t)$2 between actual and pseudo entropy (Bar et al., 2024). These designs share the premise that the proxy objective should encode a stable reference distribution rather than merely sharpen predictions.

4. Guarantees, regret, and recalibration

The theory of online distributional adaptation is heterogeneous because the underlying tasks are heterogeneous. In online label shift, the main guarantee is regret. If $p(X_t|\theta_t)$3 is convex in $p(X_t|\theta_t)$4 and $p(X_t|\theta_t)$5, then OGD with $p(X_t|\theta_t)$6 yields expected average regret bounded by $p(X_t|\theta_t)$7 (Wu et al., 2021). Under Lipschitzness and a symmetric-optimality assumption, Follow-The-History achieves a high-probability bound with an ideal-case rate $p(X_t|\theta_t)$8. OLS-OFU preserves the same online-regret convergence order as the base OLS oracle, but replaces the comparator $p(X_t|\theta_t)$9 by $f_{\theta^S}$0; whenever the self-supervised feature updates improve the feature extractor, the regret becomes strictly smaller (Wu et al., 2024).

For wrapper methods, guarantees concern adaptivity to unknown time scales. AWE maintains only $f_{\theta^S}$1 active Online Learner instances through a multi-resolution interval design, and Theorem 4 bounds instantaneous regret by the error of the best restarted instance plus an additional $f_{\theta^S}$2 term arising from adaptive cross-validation among active instances (Baby et al., 9 Apr 2025). The dynamic-regret result scales as

$f_{\theta^S}$3

where $f_{\theta^S}$4 denotes total variation within bins. This formalizes the “attention span” problem: the relevant history length changes with the shift process.

In uncertainty-sensitive prediction, the guarantees change form. RetroAdj for online conformal inference uses retrospective adjustment of leave-one-out residuals in a sliding window and achieves long-run coverage:

$f_{\theta^S}$5

when combined with any of the ACI-type updates on $f_{\theta^S}$6 (Jun et al., 6 Nov 2025). LightTune proves that, under boundedness, smoothness, and a gradient lower bound under large error, the average probability of a large error converges to zero despite $f_{\theta^S}$7 (Ali et al., 14 Apr 2026). POEM proves that the wealth process induced by its betting martingale is valid under no shift and that false alarms are controlled by Ville’s inequality; its adaptive SF-OGD update on the betting strength attains $f_{\theta^S}$8 regret (Bar et al., 2024).

These results do not imply a unified theory, but they do isolate recurring technical themes: unbiased surrogate construction without labels, adaptive selection of temporal scale, and explicit control of miscoverage, false alarms, or large-error events. A plausible implication is that online distributional adaptation is increasingly analyzed through task-specific online decision criteria rather than a single universal stability theorem.

5. Empirical evidence across domains

Empirical work shows that online distributional adaptation is useful across markedly different modalities, though the best mechanism depends on the shift structure. In online label shift on CIFAR-10 with $f_{\theta^S}$9 under constant, monotone, periodic, and exponential-periodic shifts, all online methods outperform the static base model; FTH tracks the Optimal Fixed Classifier but is conservative, while OGD with the smooth surrogate often matches or even improves upon OFC and achieves the lowest error in most scenarios (Wu et al., 2021). On ArXiv paper-category data with 23 classes and $p_S(x,y)$0, OGD (surrogate) attains the best error at approximately $p_S(x,y)$1, close to OFC at approximately $p_S(x,y)$2, with the static base at approximately $p_S(x,y)$3.

Meta-learned stream adaptation improves conventional TTA baselines in both synthetic and real shifts. On CIFAR-10-C with periodic shifts of length $p_S(x,y)$4, the reported accuracy is approximately $p_S(x,y)$5 versus $p_S(x,y)$6 for CoTTA; on Tiny ImageNet-C with $p_S(x,y)$7, the reported result is $p_S(x,y)$8 versus $p_S(x,y)$9 for the best baseline; and on CamVid video segmentation the method reaches mIoU $p_T(x,y)$0 versus $p_T(x,y)$1 for CoTTA and $p_T(x,y)$2 for Vanilla (Wu et al., 2023). These results are explicitly tied to per-sample sequential adaptation and to support-set design during meta-training.

Alternative objectives improve robustness under covariate and corruption shifts. CoMM reports lower Top-1 error than both the source model and TENT on all listed corruption and domain adaptation benchmarks: for example, CIFAR-10-C error decreases from $p_T(x,y)$3 to $p_T(x,y)$4, and ImageNet-C ResNet-50 error from $p_T(x,y)$5 to $p_T(x,y)$6 (Chuah et al., 2024). DA-TTA reports that, in non-i.i.d. streams, CIFAR-10-C error falls from $p_T(x,y)$7 to $p_T(x,y)$8, CIFAR-100-C from $p_T(x,y)$9 to $p_S$0, and ImageNet-C from $p_S$1 to $p_S$2, while also improving continual ImageNet-C streams from $p_S$3 to $p_S$4 (Wang et al., 2024).

The same general idea extends beyond image classification. In preparation-free heterogeneous cooperative perception, direct cross-agent fusion drops below the ego-only baseline, but BOLT restores compatibility through an online adaptive plugin with only $p_S$5 trainable parameters; every tested encoder pair exceeds ego-only performance, with gains up to $p_S$6 AP@50 and, in a 3-agent OPV2V test, AP@50 improving from $p_S$7 to $p_S$8 (Yang et al., 1 May 2026). In language modeling on the PG-19 test stream, dynamic evaluation improves average loss from $p_S$9 to $$\theta^t=\theta^{t-1}-\eta\nabla_\theta L(\mathcal{B}_t;\theta^{t-1})$$0 nats/token without reset and to $$\theta^t=\theta^{t-1}-\eta\nabla_\theta L(\mathcal{B}_t;\theta^{t-1})$$1 nats/token with reset per book; the cumulative-loss tipping point typically occurs after roughly $$\theta^t=\theta^{t-1}-\eta\nabla_\theta L(\mathcal{B}_t;\theta^{t-1})$$2–$$\theta^t=\theta^{t-1}-\eta\nabla_\theta L(\mathcal{B}_t;\theta^{t-1})$$3k tokens (Rannen-Triki et al., 2024). In mobile link adaptation, LightTune reduces average BLER prediction error by up to $$\theta^t=\theta^{t-1}-\eta\nabla_\theta L(\mathcal{B}_t;\theta^{t-1})$$4 and yields average throughput improvements of up to $$\theta^t=\theta^{t-1}-\eta\nabla_\theta L(\mathcal{B}_t;\theta^{t-1})$$5 compared to a conventional table-based OLLA algorithm (Ali et al., 14 Apr 2026).

6. Limitations, misconceptions, and future directions

A persistent misconception is that any online adaptation objective will improve performance so long as the model is updated on unlabeled test data. The OTTA survey reports a different picture: transformers already exhibit heightened resilience to diverse domain shifts, many OTTA methods rely on large batch sizes, and stability in optimization and resistance to perturbations are crucial during adaptation, particularly when the batch size is $$\theta^t=\theta^{t-1}-\eta\nabla_\theta L(\mathcal{B}_t;\theta^{t-1})$$6 (Wang et al., 2023). The survey’s findings diverge from existing literature precisely because ambiguous settings, outdated backbones, and inconsistent hyperparameter tuning can obscure the core challenges.

A second misconception is that entropy minimization is a universally adequate surrogate. CoMM identifies two key flaws in entropy minimization: noisy gradients near decision boundaries and reinforcement of overconfident but incorrect predictions under strong domain shift (Chuah et al., 2024). POEM reaches a related conclusion from a different angle, arguing that adaptation should match the entropy distribution to that of the source rather than simply reduce entropy (Bar et al., 2024). DA-TTA similarly argues that self-training losses can create conflicting optimization objectives in non-i.i.d. streams with prominent label shifts across batches, and therefore reverses the adaptation direction by aligning test features back to source statistics (Wang et al., 2024).

Practical limitations are equally prominent. BOLT assumes a shared BEV grid and incurs an online update cost of approximately $$\theta^t=\theta^{t-1}-\eta\nabla_\theta L(\mathcal{B}_t;\theta^{t-1})$$7–$$\theta^t=\theta^{t-1}-\eta\nabla_\theta L(\mathcal{B}_t;\theta^{t-1})$$8ms per frame, which constrains use under tight latency budgets (Yang et al., 1 May 2026). Dynamic evaluation requires backward passes at inference time and can need several thousand tokens before beating static evaluation, so its benefits depend strongly on context budget, model size, and distribution shift magnitude (Rannen-Triki et al., 2024). OLS-OFU can underperform simple OFU under very severe non-label shifts that violate the fixed $$\theta^t=\theta^{t-1}-\eta\nabla_\theta L(\mathcal{B}_t;\theta^{t-1})$$9 assumption even after self-supervised refinement (Wu et al., 2024). Meta-trained stream adaptation requires access to many simulated domains and can be compute-intensive, while its effectiveness remains tied to the quality of the self-supervised proxy (Wu et al., 2023).

Current research directions are correspondingly diverse. OTTA benchmarks emphasize realistic streaming protocols, multimodal OTTA and prompting, hot-swappable OTTA modules, and universal stability improvements across backbones and batch sizes (Wang et al., 2023). Probabilistic state-space formulations propose explicit prior tuning, transition modeling, and uncertainty propagation in parameter space (Corrales et al., 24 Jun 2026). Other work suggests retrospective recalibration for conformal prediction, lightweight forward-only updates for edge devices, and plug-and-play adaptation modules for heterogeneous multi-agent systems (Jun et al., 6 Nov 2025). Taken together, these developments indicate that online distributional adaptation is no longer a single technique but a research area organized around a common deployment constraint: the model must react to evolving distributions before labels, source data, or retraining pipelines become available.