Research Papers

Audio Large Language Models (Audio LLMs) enable human-like conversation about music, yet it is unclear if they are truly listening to the audio or just using textual reasoning, as recent benchmarks suggest. This paper investigates this issue by quantifying the contribution of each modality to a model's output. We adapt the MM-SHAP framework, a performance-agnostic score based on Shapley values that quantifies the relative contribution of each modality to a model's prediction. We evaluate two models on the MuChoMusic benchmark and find that the model with higher accuracy relies more on text to answer questions, but further inspection shows that even if the overall audio contribution is low, models can successfully localize key sound events, suggesting that audio is not entirely ignored. Our study is the first application of MM-SHAP to Audio LLMs and we hope it will serve as a foundational step for future research in explainable AI and audio.

The adoption of cyber-physical systems and jobsite intelligence that connects design models, real-time site sensing, and autonomous field operations can dramatically enhance digital management in the construction industry. This paper introduces BIM2RDT (Building Information Models to Robot-Ready Site Digital Twins), an agentic artificial intelligence (AI) framework designed to transform static Building Information Modeling (BIM) into dynamic, robot-ready digital twins (DTs) that prioritize safety during execution. The framework bridges the gap between pre-existing BIM data and real-time site conditions by integrating three key data streams: geometric and semantic information from BIM models, activity data from IoT sensor networks, and visual-spatial data collected by robots during site traversal. The methodology introduces Semantic-Gravity ICP (SG-ICP), a point cloud registration algorithm that leverages large language model (LLM) reasoning. Unlike traditional methods, SG-ICP utilizes an LLM to infer object-specific, plausible orientation priors based on BIM semantics, improving alignment accuracy by avoiding convergence on local minima. This creates a feedback loop where robot-collected data updates the DT, which in turn optimizes paths for missions. The framework employs YOLOE object detection and Shi-Tomasi corner detection to identify and track construction elements while using BIM geometry as a priori maps. The framework also integrates real-time Hand-Arm Vibration (HAV) monitoring, mapping sensor-detected safety events to the digital twin using IFC standards for intervention. Experiments demonstrate SG-ICP's superiority over standard ICP, achieving RMSE reductions of 64.3%--88.3% in alignment across scenarios with occluded features, ensuring plausible orientations. HAV integration triggers warnings upon exceeding exposure limits, enhancing compliance with ISO 5349-1.

Language Models (LMs) have significantly advanced natural language processing and enabled remarkable progress across diverse domains, yet their black-box nature raises critical concerns about the interpretability of their internal mechanisms and decision-making processes. This lack of transparency is particularly problematic for adoption in high-stakes domains, where stakeholders need to understand the rationale behind model outputs to ensure accountability. On the other hand, while explainable artificial intelligence (XAI) methods have been well studied for non-LMs, they face many limitations when applied to LMs due to their complex architectures, considerable training corpora, and broad generalization abilities. Although various surveys have examined XAI in the context of LMs, they often fail to capture the distinct challenges arising from the architectural diversity and evolving capabilities of these models. To bridge this gap, this survey presents a comprehensive review of XAI techniques with a particular emphasis on LMs, organizing them according to their underlying transformer architectures: encoder-only, decoder-only, and encoder-decoder, and analyzing how methods are adapted to each while assessing their respective strengths and limitations. Furthermore, we evaluate these techniques through the dual lenses of plausibility and faithfulness, offering a structured perspective on their effectiveness. Finally, we identify open research challenges and outline promising future directions, aiming to guide ongoing efforts toward the development of robust, transparent, and interpretable XAI methods for LMs.

This paper critically analyzes the discourse of the 'AI executive elite,' a group of highly influential individuals shaping the way AI is funded, developed, and deployed worldwide. The primary objective is to examine the presence and dynamics of the 'Techno-Supremacy Doctrine' (TSD), a term introduced in this study to describe a belief system characterized by an excessive trust in technology's alleged inherent superiority in solving complex societal problems. This study integrates quantitative heuristics with in-depth qualitative investigations. Its methodology is operationalized in a two-phase critical discourse analysis of 14 texts published by elite members between 2017 and 2025. The findings demonstrate that the elite is not a monolithic bloc but exhibits a broad spectrum of stances. The discourse is highly dynamic, showing a marked polarization and general increase in pro-TSD discourse following the launch of ChatGPT. The analysis identifies key discursive patterns, including a dominant pro-TSD narrative that combines utopian promises with claims of inevitable progress, and the common tactic of acknowledging risks only as a strategic preamble to proposing further technological solutions. This paper presents TSD as a comprehensive analytical framework and provides a 'diagnostic toolkit' for identifying its manifestations, from insidious to benign. It argues that fostering critical awareness of these discursive patterns is essential for AI practitioners, policymakers, and the public to actively navigate the future of AI.

Large Language Models have demonstrated impressive fluency across diverse tasks, yet their tendency to produce toxic content remains a critical challenge for AI safety and public trust. Existing toxicity mitigation approaches primarily manipulate individual neuron activations, but these methods suffer from instability, context dependence, and often compromise the model's core language abilities. To address these shortcomings, we investigate three key questions: the stability of neuron-level toxicity indicators, the advantages of structural (layer-wise) representations, and the interpretability of mechanisms driving toxic generation. Through extensive experiments on Jigsaw and ToxiCN datasets, we show that aggregated layer-wise features provide more robust signals than single neurons. Moreover, we observe conceptual limitations in prior works that conflate toxicity detection experts and generation experts within neuron-based interventions. To mitigate this, we propose a novel principled intervention technique, EigenShift, based on eigen-decomposition of the language model's final output layer. This method selectively targets generation-aligned components, enabling precise toxicity suppression without impairing linguistic competence. Our method requires no additional training or fine-tuning, incurs minimal computational cost, and is grounded in rigorous theoretical analysis.

The increasing use of machine learning in clinical decision support has been limited by the lack of transparency of many high-performing models. In clinical settings, predictions must be interpretable, auditable, and actionable. This study investigates Kolmogorov-Arnold Networks (KANs) as intrinsically interpretable alternatives to conventional black-box models for clinical classification of tabular health data, aiming to balance predictive performance with clinically meaningful transparency. We introduce two KAN-based models: the Logistic KAN, a flexible generalization of logistic regression, and the Kolmogorov-Arnold Additive Model (KAAM), an additive variant that yields transparent symbolic representations through feature-wise decomposability. Both models are evaluated on multiple public clinical datasets and compared with standard linear, tree-based, and neural baselines. Across all datasets, the proposed models achieve predictive performance comparable to or exceeding that of commonly used baselines while remaining fully interpretable. Logistic-KAN obtains the highest overall ranking across evaluation metrics, with a mean reciprocal rank of 0.76, indicating consistently strong performance across tasks. KAAM provides competitive accuracy while offering enhanced transparency through feature-wise decomposability, patient-level visualizations, and nearest-patient retrieval, enabling direct inspection of individual predictions. KAN-based models provide a practical and trustworthy alternative to black-box models for clinical classification, offering a strong balance between predictive performance and interpretability for clinical decision support. By enabling transparent, patient-level reasoning and clinically actionable insights, the proposed models represent a promising step toward trustworthy AI in healthcare (code: https://github.com/Patricia-A-Apellaniz/classification_with_kans).

Explainable AI for Infection Prevention and Control: Modeling CPE Acquisition and Patient Outcomes in an Irish Hospital with Transformers

Minh-Khoi Pham, Tai Tan Mai, Martin Crane +9

Carbapenemase-Producing Enterobacteriace poses a critical concern for infection prevention and control in hospitals. However, predictive modeling of previously highlighted CPE-associated risks such as readmission, mortality, and extended length of stay (LOS) remains underexplored, particularly with modern deep learning approaches. This study introduces an eXplainable AI modeling framework to investigate CPE impact on patient outcomes from Electronic Medical Records data of an Irish hospital. We analyzed an inpatient dataset from an Irish acute hospital, incorporating diagnostic codes, ward transitions, patient demographics, infection-related variables and contact network features. Several Transformer-based architectures were benchmarked alongside traditional machine learning models. Clinical outcomes were predicted, and XAI techniques were applied to interpret model decisions. Our framework successfully demonstrated the utility of Transformer-based models, with TabTransformer consistently outperforming baselines across multiple clinical prediction tasks, especially for CPE acquisition (AUROC and sensitivity). We found infection-related features, including historical hospital exposure, admission context, and network centrality measures, to be highly influential in predicting patient outcomes and CPE acquisition risk. Explainability analyses revealed that features like "Area of Residence", "Admission Ward" and prior admissions are key risk factors. Network variables like "Ward PageRank" also ranked highly, reflecting the potential value of structural exposure information. This study presents a robust and explainable AI framework for analyzing complex EMR data to identify key risk factors and predict CPE-related outcomes. Our findings underscore the superior performance of the Transformer models and highlight the importance of diverse clinical and network features.

We present an explainable AI-enhanced supervisory control framework for multi-agent robotics that combines (i) a timed-automata supervisor for safe, auditable mode switching, (ii) robust continuous control (Lyapunov-based controller for large-angle maneuver; sliding-mode controller (SMC) with boundary layers for precision and disturbance rejection), and (iii) an explainable predictor that maps mission context to gains and expected performance (energy, error). Monte Carlo-driven optimization provides the training data, enabling transparent real-time trade-offs. We validated the approach in two contrasting domains, spacecraft formation flying and autonomous underwater vehicles (AUVs). Despite different environments (gravity/actuator bias vs. hydrodynamic drag/currents), both share uncertain six degrees of freedom (6-DOF) rigid-body dynamics, relative motion, and tight tracking needs, making them representative of general robotic systems. In the space mission, the supervisory logic selects parameters that meet mission criteria. In AUV leader-follower tests, the same SMC structure maintains a fixed offset under stochastic currents with bounded steady error. In spacecraft validation, the SMC controller achieved submillimeter alignment with 21.7% lower tracking error and 81.4% lower energy consumption compared to Proportional-Derivative PD controller baselines. At the same time, in AUV tests, SMC maintained bounded errors under stochastic currents. These results highlight both the portability and the interpretability of the approach for safety-critical, resource-constrained multi-agent robotics.

Is "Knowing It's Malicious Enough?" Evaluating LLMs for Fine-Grained Malware Behavior Auditing

Xinran Zheng, Xingzhi Qian, Yiling He +2

Automated malware classifiers achieve strong detection performance, but auditing requires more than flagging a sample: analysts must explain malicious behaviors and justify them with code evidence. Traditional signature-based methods and learning-based XAI often fail to provide such support in a human-interpretable form. Large Language Models (LLMs) appear promising, yet their reliability for malware auditing remains unclear. Evaluation faces three challenges: (1) the lack of human-written behavioral ground truth; (2) real-world codebases that exceed current context limits; and (3) the lack of reliable mechanisms to verify whether generated claims are grounded in code evidence. These obstacles make benchmarking difficult and leave model capabilities and failure modes opaque. We introduce MalEval, a diagnostic framework for measuring the capability boundaries of LLMs in malware auditing. MalEval pairs real-world application codebases with expert-written audit reports to provide fine-grained behavior-level ground truth. It compresses large codebases into behavior-relevant program contexts through a context-driven intermediate representation that preserves call relations. Expert reports and model outputs are mapped, via constrained reasoning, into structured evidence chains linking code-level facts to high-level behaviors in a shared space. MalEval decomposes auditing into four stage-wise tasks, enabling each intermediate judgment to be verified under limited context windows. We evaluate seven LLMs and find that they rely on surface cues rather than verifiable evidence, struggle to compose dispersed facts into coherent attack chains, and are highly sensitive to context formulation. These findings shift attention from isolated outputs to reliable LLM and agentic workflows for malware auditing. MalEval is publicly available at https://github.com/ZhengXR930/MalEval.git

Effective model and hyperparameter selection remains a major challenge in deep learning, often requiring extensive expertise and computation. While AutoML and large language models (LLMs) promise automation, current LLM-based approaches rely on trial and error and expensive APIs, which provide limited interpretability and generalizability. We propose MetaLLMiX, a zero-shot hyperparameter optimization framework combining meta-learning, explainable AI, and efficient LLM reasoning. By leveraging historical experiment outcomes with SHAP explanations, MetaLLMiX recommends optimal hyperparameters and pretrained models without additional trials. We further employ an LLM-as-judge evaluation to control output format, accuracy, and completeness. Experiments on eight medical imaging datasets using nine open-source lightweight LLMs show that MetaLLMiX achieves competitive or superior performance to traditional HPO methods while drastically reducing computational cost. Our local deployment outperforms prior API-based approaches, achieving optimal results on 5 of 8 tasks, response time reductions of 99.6-99.9%, and the fastest training times on 6 datasets (2.4-15.7x faster), maintaining accuracy within 1-5% of best-performing baselines.

Measuring Epistemic Humility in Multimodal Large Language Models

Bingkui Tong, Jiaer Xia, Sifeng Shang +1

Hallucinations in multimodal large language models (MLLMs), where a model generates content inconsistent with the input image, pose significant risks in real-world applications. Existing hallucination benchmarks primarily test recognition, i.e., whether a model can select the correct answer among distractors. They overlook another capability that is essential for trustworthy AI: recognizing when none of the provided options is supported by the image and declining to commit to a false choice. We present HumbleBench, a hallucination benchmark that evaluates this false-option rejection behavior, a directly observable facet of epistemic humility, through multiple-choice questions that always include a "None of the above" (NOTA) option. HumbleBench is built from the Panoptic Scene Graph dataset: its object and relation annotations, together with candidate attribute cues, are used to prompt GPT-4-Turbo to generate questions, which are then manually verified, yielding 22,831 questions on object, relation, and attribute hallucinations. We evaluate 27 MLLMs, including general-purpose, reasoning-oriented, and proprietary models, and complement standard accuracy with NOTA-aware metrics, diagnostic stress tests, and a human evaluation. Even the strongest models select NOTA for only about half of the questions in which it is correct, cautious prompting shifts models toward NOTA without improving their ability to detect unsupported options, and humans reject unsupported options considerably more reliably than all evaluated models. These results show that recognition accuracy alone is insufficient for assessing hallucination robustness. Our code and dataset are publicly available at https://github.com/maifoundations/HumbleBench.

Explainable Artificial Intelligence (XAI) enhances the transparency and interpretability of AI models, addressing their inherent opacity. In cybersecurity, particularly within the Internet of Medical Things (IoMT), the black-box nature of AI-driven threat detection poses a significant challenge. Cybersecurity professionals must not only detect attacks but also understand the reasoning behind AI decisions to ensure trust and accountability. The rapid increase in cyberattacks targeting connected medical devices threatens patient safety and data privacy, necessitating advanced AI-driven solutions. This study compares two ensemble learning techniques, bagging and boosting, for cyber-attack classification in IoMT environments. We selected Random Forest for bagging and CatBoost for boosting. Random Forest helps reduce variance, while CatBoost improves bias by combining weak classifiers into a strong ensemble model, making them effective for detecting sophisticated attacks. However, their complexity often reduces transparency, making it difficult for cybersecurity professionals to interpret and trust their decisions. To address this issue, we apply XAI models to generate local and global explanations, providing insights into AI decision-making. Using techniques like SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-agnostic Explanations), we highlight feature importance to help stakeholders understand the key factors driving cyber threat detection.