A new arXiv paper proposes updating intensive care prediction models through a structurally bounded procedure that leaves the physiological encoder frozen. The authors argue this limits how much model behavior and its explanations can drift when patient data distributions change. The aim is to make adapted clinical models easier to audit after deployment.
A new arXiv paper introduces PIDS-Bench, a benchmark that evaluates prompt-injection detectors beyond aggregate F1 scores on in-distribution test data. It examines detector behavior under distribution shift, obfuscation and over-defense, with particular attention to false positives on benign inputs near the decision boundary. The authors argue that standard evaluation practices give limited visibility into how these systems behave in realistic conditions.
A new arXiv paper proposes a conformal prediction approach that draws on multiple heterogeneous data sources rather than treating them as one pool. The method exploits differences between sources through localization, aiming to keep prediction sets reliable when the test distribution departs from any single source. This targets settings where combining sources is useful but naive pooling would break coverage guarantees.
A new arXiv preprint introduces a method that organizes clinical data into hierarchy-based domains with adjustable granularity, aiming to improve model generalization across shifting patient populations. The authors argue that standard domain generalization techniques fall short in healthcare settings, where data distributions vary between patient groups. The work was posted as a replacement version on arXiv's machine learning and AI categories.
A new arXiv preprint examines how machine-learning pipelines behave when training or test data is incomplete, imbalanced, poorly labelled, or drawn from mismatched distributions. The authors survey measurement approaches and methods designed to keep models reliable under these common real-world conditions. The work is framed as an overview of challenges and remedies rather than a new model release.
Researchers present TimeWarp, a benchmark designed to reflect how websites change over time, addressing the concern that agents may not generalize from today's web to tomorrow's. It comprises three web environments that emulate an evolving web so agent performance can be measured under such shifts. The work appears as a replacement submission on arXiv.
A new arXiv paper introduces TypiCore, a method that combines active learning queries with class-incremental learning for time series data. The approach targets settings such as healthcare and manufacturing, where models must adapt to distribution shifts as new classes arrive over time. The work is posted as a cross-listing replacement on arXiv and has not yet undergone peer review.
This paper examines how to estimate density ratios and perform importance-weighted regression when the output distribution shifts between training and test data while the conditional input distribution given outputs stays the same. The authors derive optimal estimation approaches for this continuous-output target shift setting. The work falls in the statistical machine learning area of distribution shift and covariate/label shift correction.
A new arXiv paper argues that comparing large language models with classical machine learning on network intrusion detection only within a single dataset gives an incomplete picture. The authors evaluate XGBoost and a RoBERTa-LoRA model under distribution shift and adversarial evasion to probe how each approach holds up outside the usual same-dataset setup. The work highlights robustness gaps that standard benchmarks tend to miss.
A new arXiv preprint addresses unlabeled-unlabeled (UU) learning, a setting where a binary classifier is trained from two unlabeled datasets that have differing class priors. The authors introduce importance weighting to handle distribution shift in this framework, which generalizes approaches such as positive-unlabeled learning. The work is listed as a cross-submission in the cs.AI category.
A new arXiv paper proposes a general approach to measuring how covariate and concept shifts affect machine learning generalization. The authors argue that existing learning bound theory covers only narrow, idealized settings and cannot be estimated from data. Their framework aims to make distribution shift analysis broadly applicable and computable from samples.
A new machine learning paper tackles the gap between training conditions and real-world deployment for sensor-based AI, where devices, personnel, and time periods all differ from the original training setup. The authors propose a staged evaluation methodology that quantifies uncertainty and captures performance degradation that conventional random train-test splits can hide. The work draws on data collected from multiple devices and subjects over nearly three years in an underground environment.
A new arXiv preprint examines how to choose a deferral policy for a fixed classifier, where uncertain cases are routed to human reviewers. The authors analyze how miscalibrated confidence scores, unequal error costs, fallible reviewers, and deployment-time distribution shift interact, using an environmental AI application as a real-world case study. The work offers practical guidance for deciding when automated predictions should be handed off rather than trusted outright.
A new arXiv paper presents a method that distills image prototypes to guide test-time adaptation of models facing distribution shifts. The technique addresses two common failure modes in this setting: error buildup from unreliable pseudo-labels and degradation of knowledge learned during pretraining. By anchoring adaptation to distilled prototypes, the authors aim to make inference-time model updates more stable.