online adaptation

To address degraded model adaptability in online multistep time-series forecasting—caused by data distribution drift and delayed ground-truth feedback—this paper proposes ADAPT-Z. Methodologically, ADAPT-Z abandons conventional parameter fine-tuning and instead models the dynamics of latent factors. It introduces an adapter module that fuses current features with historical gradient information within a learned Z-space, enabling persistent tracking and incremental self-adaptation of feature representations. This design mitigates gradient mismatch induced by label delay and enhances robustness to non-stationary data. Empirical evaluation across multiple benchmark datasets demonstrates that ADAPT-Z significantly outperforms static baselines and state-of-the-art online learning methods, achieving superior generalization and sustained adaptive capability under streaming conditions.

To ensure safety in high-risk AI systems, continuous monitoring for abrupt distributional shifts—such as concept drift, covariate shift, and out-of-support shifts—is essential post-deployment. This paper proposes the Weighted Conformal Test Martingale (WCTM), a generalized nonparametric online changepoint detection framework. WCTM is the first method to achieve *anytime-valid* detection of arbitrary distributional shifts while rigorously controlling the false alarm rate; it further supports online adaptation to mild covariate shifts. Its theoretical foundation integrates weighted conformal prediction, anytime-valid inference, and online martingale construction, unifying the modeling of covariate, concept, and support-set shifts. Evaluated on multiple real-world datasets, WCTM significantly outperforms state-of-the-art methods, achieving superior trade-offs between detection sensitivity and false positive rate, and demonstrating distinct responsiveness to both adaptive and non-adaptive shifts.

This work addresses the challenge of natural distribution shift—gradually evolving over time—in model deployment, proposing a label-free online test-time adaptation (TTA) method. Unlike prevailing TTA approaches designed for synthetic corruptions, our method is the first to integrate stochastic state-space models (SSMs) into the TTA framework. By performing latent-variable inference, it explicitly models time-varying dynamics in feature representations, enabling unsupervised, dynamic class-prototype learning and adaptive classifier-head updating. Evaluated on realistic temporal distribution-shift benchmarks, our approach significantly outperforms existing TTA methods, particularly under small-batch inference and label-shift conditions, demonstrating superior robustness and performance gains. The method establishes a novel paradigm for open-world continual learning under non-stationary environments.

Distribution shift in tabular data lacks empirical grounding, and existing robust learning methods—such as distributionally robust optimization (DRO)—frequently fail in real-world settings. Prior work predominantly assumes covariate shift (X-shift), whereas empirical analysis reveals that label-conditional shift (Y|X-shift) is both more prevalent and dominant in practice. Method: The authors construct a comprehensive benchmark platform spanning five real-world tabular datasets and 60,000 experimental configurations, enabling large-scale ablation studies on robustness. Contribution/Results: (1) Theory-driven methods like DRO offer no consistent advantage over empirical risk minimization (ERM); (2) implementation details—including model selection, hyperparameter tuning, and imbalance handling—exert far greater influence on robustness than the design of ambiguity sets; (3) a data-driven, inductive paradigm should supplant prior structural assumptions. These findings challenge the conventional paradigm in distributionally robust learning and provide reproducible, empirically grounded guidance for co-optimizing algorithms and data.

Real-world tabular data often suffer from distribution shifts—particularly label distribution shift (LDS)—leading to substantial degradation in model performance; existing test-time adaptation (TTA) methods either neglect LDS characteristics or are constrained by architectural assumptions, limiting applicability in privacy-sensitive settings. This paper proposes the first source-free TTA framework for tabular data, introducing a novel two-stage decoupled mechanism: an offset-aware uncertainty calibrator jointly optimizes with a label distribution processor to systematically model and mitigate LDS for the first time. The method integrates uncertainty calibration, label distribution estimation and reweighting, theory-guided robustness analysis, and source-free online adaptation. Evaluated on diverse LDS benchmarks—including HELOC—it achieves up to 16% accuracy improvement over state-of-the-art TTA approaches, demonstrating strong model-agnosticism and generalization capability.

This work addresses the degraded out-of-distribution (OOD) generalization performance of electrocardiogram (ECG) time series models under limited labeled data and distribution shifts—such as those arising from varying sensors, populations, or recording conditions—by proposing a fine-tuning framework that integrates data uncertainty estimation, low-rank adaptation (LoRA), and adaptive hyperparameter optimization. The method uniquely leverages data uncertainty to explicitly quantify the degree of distribution shift in time series tasks and dynamically adjusts the fine-tuning strategy accordingly. Experimental results demonstrate that the proposed framework consistently outperforms existing approaches across diverse OOD scenarios, achieving up to a 7% improvement in accuracy and a 12.9% gain in precision, while maintaining robust performance even under severe distribution shifts.

This work addresses the challenge of distribution shift between training and test phases by proposing the first online test-time adaptation framework based on state space modeling. The method learns an initial model from labeled data during training and dynamically updates its parameters using unlabeled test data at inference time. It unifies parameter learning, temporal evolution, prior refinement, and prediction within a single probabilistic state space architecture. This formulation enables recursive online parameter updates and principled uncertainty quantification, yielding a general and robust adaptation mechanism. The approach provides both theoretical grounding and an effective solution for online prediction under distributional shifts.

This study addresses the challenge of population distribution shift that arises when external control data are incorporated into randomized controlled trials due to cost constraints, which violates the conventional exchangeability assumption and biases causal effect estimation. To overcome this limitation, the authors propose a distribution-shift-aware semiparametric framework that explicitly models the distributional discrepancy between trial participants and external controls. By integrating calibration equations to adjust the efficient influence function and employing an adaptive shrinkage strategy, the method constructs an augmented estimator that maintains consistency while achieving higher statistical efficiency than estimators relying solely on trial data. Both theoretical analysis and empirical evaluations demonstrate that the proposed approach substantially improves estimation efficiency across synthetic and real-world scenarios, effectively relaxing the stringent exchangeability requirement.

Clinical risk prediction models often suffer performance degradation due to temporal distributional shifts in patient populations—such as those induced by the COVID-19 pandemic or electronic health record (EHR) system upgrades—particularly during transitional periods involving heterogeneous cohorts. To address this, we propose a lightweight online transfer learning framework that dynamically adapts models at the single-visit level without full retraining. Our approach uniquely integrates domain discrimination, adversarial feature alignment, and Bayesian calibration to enable incremental, real-time adaptation. Evaluated on real-world emergency department data spanning the pre- and early-COVID transition period, our method improves inpatient admission risk prediction AUC by 3.2 percentage points and reduces Brier score by 18.7% over static baseline models. The framework ensures computational efficiency, robustness to distributional shifts, and model interpretability, establishing a novel paradigm for continual adaptation of clinical decision support systems.

This work addresses the significant performance degradation of EEG foundation models under real-world distribution shifts across devices, populations, and recording scenarios. To this end, we present NeuroAdapt-Bench, the first benchmark for test-time adaptation (TTA) tailored to EEG foundation models, enabling systematic evaluation of representative TTA approaches—including both gradient-based and optimization-free strategies—across multiple pretrained models, downstream tasks, and heterogeneous datasets encompassing conventional, out-of-distribution, and extreme modalities such as Ear-EEG. Our experiments reveal that standard TTA methods, particularly gradient-based ones, often yield unstable or even detrimental performance on EEG tasks, whereas optimization-free methods consistently deliver reliable improvements, highlighting their practical promise for neural signal processing in clinical settings.