The increasing adoption of foundation models as agents across diverse domains necessitates a robust evaluation framework. Current methods, such as LLM-as-a-Judge, focus only on final outputs, overlooking the step-by-step reasoning that drives agentic decision-making. Meanwhile, existing Agent-as-a-Judge systems, where one agent evaluates another's task completion, are typically designed for narrow, domain-specific settings. To address this gap, we propose a generalizable, modular framework for evaluating agent task completion independent of the task domain. The framework emulates human-like evaluation by decomposing tasks into sub-tasks and validating each step using available information, such as the agent's output and reasoning. Each module contributes to a specific aspect of the evaluation process, and their outputs are aggregated to produce a final verdict on task completion. We validate our framework by evaluating the Magentic-One Actor Agent on two benchmarks, GAIA and BigCodeBench. Our Judge Agent predicts task success with closer agreement to human evaluations, achieving 4.76% and 10.52% higher alignment accuracy, respectively, compared to the GPT-4o based LLM-as-a-Judge baseline. This demonstrates the potential of our proposed general-purpose evaluation framework.

This work contributes to a real-time data-driven predictive maintenance solution for Intelligent Transportation Systems. The proposed method implements a processing pipeline comprised of sample pre-processing, incremental classification with Machine Learning models, and outcome explanation. This novel online processing pipeline has two main highlights: (i) a dedicated sample pre-processing module, which builds statistical and frequency-related features on the fly, and (ii) an explainability module. This work is the first to perform online fault prediction with natural language and visual explainability. The experiments were performed with the MetroPT data set from the metro operator of Porto, Portugal. The results are above 98 % for F-measure and 99 % for accuracy. In the context of railway predictive maintenance, achieving these high values is crucial due to the practical and operational implications of accurate failure prediction. In the specific case of a high F-measure, this ensures that the system maintains an optimal balance between detecting the highest possible number of real faults and minimizing false alarms, which is crucial for maximizing service availability. Furthermore, the accuracy obtained enables reliability, directly impacting cost reduction and increased safety. The analysis demonstrates that the pipeline maintains high performance even in the presence of class imbalance and noise, and its explanations effectively reflect the decision-making process. These findings validate the methodological soundness of the approach and confirm its practical applicability for supporting proactive maintenance decisions in real-world railway operations. Therefore, by identifying the early signs of failure, this pipeline enables decision-makers to understand the underlying problems and act accordingly swiftly.

Human teaching effort is a significant bottleneck for the broader applicability of interactive imitation learning. To reduce the number of required queries, existing methods employ active learning to query the human teacher only in uncertain, risky, or novel situations. However, during these queries, the novice's planned actions are not utilized despite containing valuable information, such as the novice's capabilities, as well as corresponding uncertainty levels. To this end, we allow the novice to say: "I plan to do this, but I am uncertain." We introduce the Active Skill-level Data Aggregation (ASkDAgger) framework, which leverages teacher feedback on the novice plan in three key ways: (1) S-Aware Gating (SAG): Adjusts the gating threshold to track sensitivity, specificity, or a minimum success rate; (2) Foresight Interactive Experience Replay (FIER), which recasts valid and relabeled novice action plans into demonstrations; and (3) Prioritized Interactive Experience Replay (PIER), which prioritizes replay based on uncertainty, novice success, and demonstration age. Together, these components balance query frequency with failure incidence, reduce the number of required demonstration annotations, improve generalization, and speed up adaptation to changing domains. We validate the effectiveness of ASkDAgger through language-conditioned manipulation tasks in both simulation and real-world environments. Code, data, and videos are available at https://askdagger.github.io.

The increasing deployment of intelligent agents in digital ecosystems, such as social media platforms, has raised significant concerns about traceability and accountability, particularly in cybersecurity and digital content protection. Traditional large language model (LLM) watermarking techniques, which rely on token-level manipulations, are ill-suited for agents due to the challenges of behavior tokenization and information loss during behavior-to-action translation. To address these issues, we propose Agent Guide, a novel behavioral watermarking framework that embeds watermarks by guiding the agent's high-level decisions (behavior) through probability biases, while preserving the naturalness of specific executions (action). Our approach decouples agent behavior into two levels, behavior (e.g., choosing to bookmark) and action (e.g., bookmarking with specific tags), and applies watermark-guided biases to the behavior probability distribution. We employ a z-statistic-based statistical analysis to detect the watermark, ensuring reliable extraction over multiple rounds. Experiments in a social media scenario with diverse agent profiles demonstrate that Agent Guide achieves effective watermark detection with a low false positive rate. Our framework provides a practical and robust solution for agent watermarking, with applications in identifying malicious agents and protecting proprietary agent systems.

The rapid proliferation of online multimedia content has intensified the spread of hate speech, presenting critical societal and regulatory challenges. While recent work has advanced multimodal hateful video detection, most approaches rely on coarse, video-level annotations that overlook the temporal granularity of hateful content. This introduces substantial label noise, as videos annotated as hateful often contain long non-hateful segments. In this paper, we investigate the impact of such label ambiguity through a fine-grained approach. Specifically, we trim hateful videos from the HateMM and MultiHateClip English datasets using annotated timestamps to isolate explicitly hateful segments. We then conduct an exploratory analysis of these trimmed segments to examine the distribution and characteristics of both hateful and non-hateful content. This analysis highlights the degree of semantic overlap and the confusion introduced by coarse, video-level annotations. Finally, controlled experiments demonstrated that time-stamp noise fundamentally alters model decision boundaries and weakens classification confidence, highlighting the inherent context dependency and temporal continuity of hate speech expression. Our findings provide new insights into the temporal dynamics of multimodal hateful videos and highlight the need for temporally aware models and benchmarks for improved robustness and interpretability. Code and data are available at https://github.com/Multimodal-Intelligence-Lab-MIL/HatefulVideoLabelNoise.

--Reliable evaluation of machine learning models for neonatal seizure detection is critical for clinical adoption. Current practices often rely on inconsistent and biased metrics, hindering model comparability and interpretability. Expert-level claims about AI performance are frequently made without rigorous validation, raising concerns about their reliability. This study aims to systematically evaluate common performance metrics and propose best practices tailored to the specific challenges of neonatal seizure detection. Using real and synthetic seizure annotations, we assessed standard performance metrics, consensus strategies, and human-expert level equivalence tests under varying class imbalance, inter-rater agreement, and number of raters. Matthews and Pearson's correlation coefficients outperformed the area under the curve in reflecting performance under class imbalance. Consensus types are sensitive to the number of raters and agreement level among them. Among human-expert level equivalence tests, the multi-rater T uring test using Fleiss' κ best captured expert-level AI performance. We recommend reporting: (1) at least one balanced metric, (2) Sensitivity, specificity, PPV and NPV, (3) Multi-rater T uring test results using Fleiss' κ, and (4) All the above on held-out validation set. This proposed framework provides an important prerequisite to clinical validation by enabling a thorough and honest appraisal of AI methods for neonatal seizure detection. Neonatal seizures represent a common neurological emergency in neonatal intensive care units (NICUs) [1], [2], most commonly caused by hypoxic-ischemic encephalopathy (HIE) and cerebrovascular injury [1].

We developed a ResNet-based human activity recognition (HAR) model with minimal overhead to detect gait versus non-gait activities and everyday activities (walking, running, stairs, standing, sitting, lying, sit-to-stand transitions). The model was trained and evaluated using smartphone sensor data from adult healthy controls (HC) and people with multiple sclerosis (PwMS) with Expanded Disability Status Scale (EDSS) scores between 0.0-6.5. Datasets included the GaitLab study (ISRCTN15993728), an internal Roche dataset, and publicly available data sources (training only). Data from 34 HC and 68 PwMS (mean [SD] EDSS: 4.7 [1.5]) were included in the evaluation. The HAR model showed 98.4% and 99.6% accuracy in detecting gait versus non-gait activities in the GaitLab and Roche datasets, respectively, similar to a comparative state-of-the-art ResNet model (99.3% and 99.4%). For everyday activities, the proposed model not only demonstrated higher accuracy than the state-of-the-art model (96.2% vs 91.9%; internal Roche dataset) but also maintained high performance across 9 smartphone wear locations (handbag, shopping bag, crossbody bag, backpack, hoodie pocket, coat/jacket pocket, hand, neck, belt), outperforming the state-of-the-art model by 2.8% - 9.0%. In conclusion, the proposed HAR model accurately detects everyday activities and shows high robustness to various smartphone wear locations, demonstrating its practical applicability.

Multimodal sentiment analysis (MSA) aims to understand human emotions by integrating information from multiple modalities, such as text, audio, and visual data. However, existing methods often suffer from spurious correlations both within and across modalities, leading models to rely on statistical shortcuts rather than true causal relationships, thereby undermining generalization. To mitigate this issue, we propose a Multi-relational Multimodal Causal Intervention (MMCI) model, which leverages the backdoor adjustment from causal theory to address the confounding effects of such shortcuts. Specifically, we first model the multimodal inputs as a multi-relational graph to explicitly capture intra- and inter-modal dependencies. Then, we apply an attention mechanism to separately estimate and disentangle the causal features and shortcut features corresponding to these intra- and inter-modal relations. Finally, by applying the backdoor adjustment, we stratify the shortcut features and dynamically combine them with the causal features to encourage MMCI to produce stable predictions under distribution shifts. Extensive experiments on several standard MSA datasets and out-of-distribution (OOD) test sets demonstrate that our method effectively suppresses biases and improves performance.

Trustworthy interpretation of deep learning models is critical for neuroimaging applications, yet commonly used Explainable AI (XAI) methods lack rigorous validation, risking misinterpretation. We performed the first large-scale, systematic comparison of XAI methods on ~45,000 structural brain MRIs using a novel XAI validation framework. This framework establishes verifiable ground truth by constructing prediction tasks with known signal sources - from localized anatomical features to subject-specific clinical lesions - without artificially altering input images. Our analysis reveals systematic failures in two of the most widely used methods: GradCAM consistently failed to localize predictive features, while Layer-wise Relevance Propagation generated extensive, artifactual explanations that suggest incompatibility with neuroimaging data characteristics. Our results indicate that these failures stem from a domain mismatch, where methods with design principles tailored to natural images require substantial adaptation for neuroimaging data. In contrast, the simpler, gradient-based method SmoothGrad, which makes fewer assumptions about data structure, proved consistently accurate, suggesting its conceptual simplicity makes it more robust to this domain shift. These findings highlight the need for domain-specific adaptation and validation of XAI methods, suggest that interpretations from prior neuroimaging studies using standard XAI methodology warrant re-evaluation, and provide urgent guidance for practical application of XAI in neuroimaging.

Existing work has shown that cross-validation (CV) can be used to provide an asymptotic confidence interval for the test error of a stable machine learning algorithm, and existing stability results for many popular algorithms can be applied to derive positive instances where such confidence intervals will be valid. However, in the common setting where CV is used to compare two algorithms, it becomes necessary to consider a notion of relative stability which cannot easily be derived from existing stability results, even for simple algorithms. To better understand relative stability and when CV provides valid confidence intervals for the test error difference of two algorithms, we study the soft-thresholded least squares algorithm, a close cousin of the Lasso. We prove that while stability holds when assessing the individual test error of this algorithm, relative stability fails to hold when comparing the test error of two such algorithms, even in a sparse low-dimensional linear model setting. Additionally, we empirically confirm the invalidity of CV confidence intervals for the test error difference when either soft-thresholding or the Lasso is used. In short, caution is needed when quantifying the uncertainty of CV estimates of the performance difference of two machine learning algorithms, even when both algorithms are individually stable.