Real-world adoption of closed-loop insulin delivery systems (CLIDS) in type 1 diabetes remains low, driven not by technical failure, but by diverse behavioral, psychosocial, and social barriers. We introduce ChatCLIDS, the first benchmark to rigorously evaluate LLM–driven persuasive dialogue for health behavior change. Our framework features a library of expert-validated virtual patients, each with clinically grounded, heterogeneous profiles and realistic adoption barriers, and simulates multi-turn interactions with nurse agents equipped with a diverse set of evidence-based persuasive strategies. ChatCLIDS uniquely supports longitudinal counseling and adversarial social influence scenarios, enabling robust, multi-dimensional evaluation. Our findings reveal that while larger and more reflective LLMs adapt strategies over time, all models struggle to overcome resistance, especially under realistic social pressure. These results highlight critical limitations of current LLMs for behavior change, and offer a high-fidelity, scalable testbed for advancing trustworthy persuasive AI in healthcare and beyond.

Class incremental learning (CIL) requires an agent to learn distinct tasks consecutively with knowledge retention against forgetting. Problems impeding the practice of CIL methods are twofold: (1) prompt update on non-i.i.d batch streams without boundary, namely the harsher online task-free CIL (OTCIL) scenario; (2) CIL methods suffer from heavy forgetting on learning long task streams, as shown in Fig. 1(a). To achieve efficient decision-making, the ensemble deep random vector functional link network (edRVFL) with forward regularization (-F) is proposed to replace the canonical Ridge (-R), reducing more regrets during OTCIL. Considering continuous distribution drifting on long stream, we further propose edRVFL-kF to adjust the intervention intensity of forward knowledge and derive incremental updates. edRVFL-kF can effectively avoid replay, retraining, and catastrophic forgetting while achieving lower regret over -R. Moreover, to improve robustness on non-i.i.d stream and eliminate intractable tuning of -kF, we rebuild with online Bayesian learning and propose the plug-and-play edRVFL-kF-Bayes, enabling all hard ks in multiple sub-learners to self-adapt to ever-changing distribution and optimization in OTCIL. Experiments were conducted on image datasets, including multiple evaluations, ablation tests, estimated forward, and compatibility studies, which distinctly validate the efficacy of edRVFL-kF-Bayes.

Online learning of deep neural networks faces challenges such as delayed non-incremental updating, increasing consumption, retrospective retraining, and catastrophic forgetting. To alleviate these drawbacks and achieve progressive immediate decision-making, we propose a novel Incremental Online Learning (IOL) framework of Randomized Neural Networks (Randomized NN), facilitating continuous improvements and analytics to Randomized NN performance in online scenarios. Within the framework, we further formulate IOL with ridge regularization (-R) and IOL with forward regularization (-F), both avoiding retrospective retraining and catastrophic forgetting. Moreover, the incremental algorithms for -R/-F on non-stationary batch stream are derived, featuring recursive weight updates and variable learning rates. Compared to -R, we recommend -F which improves learning performance using future unlabeled observations while further reducing online regrets to offline global experts. Additionally, we conduct a detailed analysis and theoretically derive relative cumulative regret bounds of the Randomized NN learners for -R/-F under adversarial assumptions via a novel methodology and present several corollaries, from which we observed the superiority in online learning acceleration and declined regret bounds of employing -F in IOL. Finally, our proposed methods were rigorously examined across diverse tasks, from simulation, regression, and classification tasks, to long-term time-series forecasting (LTSF) and continual learning (CL) fields, which distinctly validated the efficacy of the IOL frameworks and the advantages of forward regularization.

Forward regularization (-F) with unsupervised knowledge was advocated to replace canonical Ridge regularization (-R) in online linear learners, as it achieved a lower relative regret boundary. However, we observe that -F cannot perform as expected in practice, even possibly losing to -R for online tasks. We identify two main causes for this: (1) inappropriate intervened regularization, and (2) non-i.i.d. nature and data distribution changes in online learning (OL), both of which result in unstable posterior distribution and optima offset of the learner. To improve these, we first introduce the adjustable forward regularization (- k F), a more general -F with controllable knowledge intervention. We also derive - k F’s incremental updates with variable learning rate, and study relative regret and boundary in OL. Inspired by the regret analysis, to curb unstable penalties, we further propose - k F-Bayes style with k synchronously self-adapted to revise the intractable tuning of - k F by considering parametric posterior distribution changes in non-i.i.d. online data streams. Additionally, we integrate the - k F and - k F-Bayes into a multi-layer ensemble deep random vector functional link (edRVFL) and present two practical algorithms for batch learning, avoiding past replay and catastrophic forgetting. In experiments, we conducted tests on numerical simulation, tabular, and image datasets, where - k F-Bayes surpassed traditional -R and -F, highlighting the efficacy of ready-to-work - k F-Bayes and the great potentials of edRVFL- k F-Bayes in OL and continual learning (CL) scenarios. • - k F provides a more flexible unsupervised knowledge intervention for online learners. • We derive - k F’s incremental updates and study relative regret in OL. • We propose - k F-Bayes to consider parametric posterior changes in non-i.i.d. streams. • We integrate the - k F and - k F-Bayes into edRVFL and present two algorithms for CL.

, which leverages information retrieval specifically for generated sub-questions and re-answers these sub-questions with the relevant contextual information. Additionally, a Retrieval-Augmented Factuality Scorer is proposed to replace the original discriminator, prioritizing reasoning paths that meet high standards of factuality. Experimental results with LLaMA 3.1 show that RARE enables open-source LLMs to achieve competitive performance with top closed-source models like GPT-4 and GPT-4o. This research establishes RARE as a scalable solution for improving LLMs in domains where logical coherence and factual integrity are critical.

Large language models (LLMs) have demonstrated promising performance on medical benchmarks; however, their ability to perform medical calculations, a crucial aspect of clinical decision-making, remains underexplored and poorly evaluated. Existing benchmarks often assess only the final answer with a wide numerical tolerance, overlooking systematic reasoning failures and potentially causing serious clinical misjudgments. In this work, we revisit medical calculation evaluation with a stronger focus on clinical trustworthiness. First, we clean and restructure the MedCalc-Bench dataset and propose a new step-by-step evaluation pipeline that independently assesses formula selection, entity extraction, and arithmetic computation. Under this granular framework, the accuracy of GPT-4o drops from 62.7% to 43.6%, revealing errors masked by prior evaluations. Second, we introduce an automatic error analysis framework that generates structured attribution for each failure mode. Human evaluation confirms its alignment with expert judgment, enabling scalable and explainable diagnostics. Finally, we propose a modular agentic pipeline, MedRaC, that combines retrieval-augmented generation and Python-based code execution. Without any fine-tuning, MedRaC improves the accuracy of different LLMs from 16.35% up to 53.19%. Our work highlights the limitations of current benchmark practices and proposes a more clinically faithful methodology. By enabling transparent and transferable reasoning evaluation, we move closer to making LLM-based systems trustworthy for real-world medical applications.

Large language models (LLMs) deliver strong performance across diverse tasks but their growing computational demands make single-device inference increasingly infeasible. Distributed inference on heterogeneous accelerators offers a scalable solution, yet existing partitioning strategies often depend on oversimplified latency models that fail to capture runtime variability caused by device diversity and fluctuating input lengths. We present an adaptive framework that jointly predicts submodel latency and optimally partitions LLMs to minimize end-to-end inference time in heterogeneous clusters. A multilayer perceptron (MLP) regression model integrates multidimensional features—such as submodel masks, device characteristics, token counts, and computational demand—and employs lightweight adapter layers for rapid fine-tuning to unseen environments. Using these latency predictions, we formulate pipeline-aware model partitioning as a dynamic programming problem, reducing search complexity from exponential to polynomial. Experiments on GPT2 show that our approach achieves high prediction accuracy (91.8% within 20% error) and improves adaptability, raising accuracy from 20.3% to 76.0% with only 50 fine-tuning samples. Overall, the proposed method reduces inference latency from 234.70 ms to 130.45 ms, significantly outperforming baseline loadbalancing methods.

Causal inference and model interpretability are gaining increasing attention, particularly in the biomedical domain. Despite recent advance, decorrelating features in nonlinear environments with human-interpretable representations remains underexplored. In this study, we introduce a novel method called causal rule generation with target trial emulation framework (CRT RE), which applies randomize trial design principles to estimate the causal effect of association rules. We then incorporate such association rules for the downstream applications such as prediction of disease onsets. Extensive experiments on six healthcare datasets, including synthetic data, real-world disease collections, and MIMIC-III/IV, demonstrate the model’s superior performance. Specifically, our method achieved a β error of 0.907, outperforming DWR (1.024) and SVM (1.141). On real-world datasets, our model achieved accuracies of 0.789, 0.920, and 0.300 for Esophageal Cancer, Heart Disease, and Cauda Equina Syndrome prediction task, respectively, consistently surpassing baseline models. On the ICD code prediction tasks, it achieved AUC Macro scores of 92.8 on MIMIC-III and 96.7 on MIMIC-IV, outperforming the state-of-the-art models KEPT and MSMN. Expert evaluations further validate the model’s effectiveness, causality, and interpretability.