Research Papers

Large language models (LLMs) deployed as agents introduce significant safety risks in clinical settings due to their potential for error and single points of failure. We introduce Tiered Agentic Oversight (TAO), a hierarchical multi-agent system that enhances AI safety through layered, automated supervision. Inspired by clinical hierarchies (e.g., nurse-physician-specialist) in hospital, TAO routes tasks to specialized agents based on complexity, creating a robust safety framework through automated inter- and intra-tier communication and role-playing. Crucially, this hierarchical structure functions as an effective error-correction mechanism, absorbing up to 24% of individual agent errors before they can compound. Our experiments reveal TAO outperforms single-agent and other multi-agent systems on 4 out of 5 healthcare safety benchmarks, with up to an 8.2% improvement. Ablation studies confirm key design principles of the system: (i) its adaptive architecture is over 3% safer than static, single-tier configurations, and (ii) its lower tiers are indispensable, as their removal causes the most significant degradation in overall safety. Finally, we validated the system's synergy with human doctors in a user study where a physician, acting as the highest tier agent, provided corrective feedback that improved medical triage accuracy from 40% to 60%. Project Page: https://tiered-agentic-oversight.github.io/

Artificial Intelligence (AI) is rapidly embedded in critical decision-making systems, however their foundational ``black-box'' models require eXplainable AI (XAI) solutions to enhance transparency, which are mostly oriented to experts, making no sense to non-experts. Alarming evidence about AI's unprecedented human values risks brings forward the imperative need for transparent human-centered XAI solutions. In this work, we introduce a domain-, model-, explanation-agnostic, generalizable and reproducible framework that ensures both transparency and human-centered explanations tailored to the needs of both experts and non-experts. The framework leverages Large Language Models (LLMs) and employs in-context learning to convey domain- and explainability-relevant contextual knowledge into LLMs. Through its structured prompt and system setting, our framework encapsulates in one response explanations understandable by non-experts and technical information to experts, all grounded in domain and explainability principles. To demonstrate the effectiveness of our framework, we establish a ground-truth contextual ``thesaurus'' through a rigorous benchmarking with over 40 data, model, and XAI combinations for an explainable clustering analysis of a well-being scenario. Through a comprehensive quality and human-friendliness evaluation of our framework's explanations, we prove high content quality through strong correlations with ground-truth explanations (Spearman rank correlation=0.92) and improved interpretability and human-friendliness to non-experts through a user study (N=56). Our overall evaluation confirms trust in LLMs as HCXAI enablers, as our framework bridges the above Gaps by delivering (i) high-quality technical explanations aligned with foundational XAI methods and (ii) clear, efficient, and interpretable human-centered explanations for non-experts.

The flourishing of video generation technologies has endangered the credibility of real-world information and intensified the demand for AI-generated video detectors. Despite some progress, the lack of high-quality real-world datasets hinders the development of trustworthy detectors. In this paper, we propose GenWorld, a large-scale, high-quality, and real-world simulation dataset for AI-generated video detection. GenWorld features the following characteristics: (1) Real-world Simulation: GenWorld focuses on videos that replicate real-world scenarios, which have a significant impact due to their realism and potential influence; (2) High Quality: GenWorld employs multiple state-of-the-art video generation models to provide realistic and high-quality forged videos; (3) Cross-prompt Diversity: GenWorld includes videos generated from diverse generators and various prompt modalities (e.g., text, image, video), offering the potential to learn more generalizable forensic features. We analyze existing methods and find they fail to detect high-quality videos generated by world models (i.e., Cosmos), revealing potential drawbacks of ignoring real-world clues. To address this, we propose a simple yet effective model, SpannDetector, to leverage multi-view consistency as a strong criterion for real-world AI-generated video detection. Experiments show that our method achieves superior results, highlighting a promising direction for explainable AI-generated video detection based on physical plausibility. We believe that GenWorld will advance the field of AI-generated video detection. Project Page: https://chen-wl20.github.io/GenWorld

Advanced Persistent Threats (APTs) represent a sophisticated and persistent cy-bersecurity challenge, characterized by stealthy, multi-phase, and targeted attacks aimed at compromising information systems over an extended period. Develop-ing an effective Intrusion Detection System (IDS) capable of detecting APTs at different phases relies on selecting network traffic features. However, not all of these features are directly related to the phases of APTs. Some network traffic features may be unrelated or have limited relevance to identifying malicious ac-tivity. Therefore, it is important to carefully select and analyze the most relevant features to improve the IDS performance. This work proposes a feature selection and classification model that integrates two prominent machine learning algo-rithms: SHapley Additive exPlanations (SHAP) and Extreme Gradient Boosting (XGBoost). The aim is to develop lightweight IDS based on a selected minimum number of influential features for detecting APTs at various phases. The pro-posed method also specifies the relevant features for each phase of APTs inde-pendently. Extensive experimental results on the SCVIC-APT-2021 dataset indi-cated that our proposed approach has improved performance compared to other standard techniques. Specifically, both the macro-average F1-score and recall reached 94% and 93 %, respectively, while reducing the complexity of the detec-tion model by selecting only 12 features out of 77.

Financial literacy is increasingly dependent on the ability to interpret complex financial data and utilize advanced forecasting tools. In this context, this study proposes a novel approach that combines transformer-based time series models with explainable artificial intelligence (XAI) to enhance the interpretability and accuracy of stock price predictions. The analysis focuses on the daily stock prices of the five highest-volume banks listed in the BIST100 index, along with XBANK and XU100 indices, covering the period from January 2015 to March 2025. Models including DLinear, LTSNet, Vanilla Transformer, and Time Series Transformer are employed, with input features enriched by technical indicators. SHAP and LIME techniques are used to provide transparency into the influence of individual features on model outputs. The results demonstrate the strong predictive capabilities of transformer models and highlight the potential of interpretable machine learning to empower individuals in making informed investment decisions and actively engaging in financial markets.

OMNIGUARD: An Efficient Approach for AI Safety Moderation Across Languages and Modalities

Sahil Verma, Keegan Hines, Jeff Bilmes +4

The emerging capabilities of large language models (LLMs) have sparked concerns about their immediate potential for harmful misuse. The core approach to mitigate these concerns is the detection of harmful queries to the model. Current detection approaches are fallible, and are particularly susceptible to attacks that exploit mismatched generalization of model capabilities (e.g., prompts in low-resource languages or prompts provided in non-text modalities such as image and audio). To tackle this challenge, we propose Omniguard, an approach for detecting harmful prompts across languages and modalities. Our approach (i) identifies internal representations of an LLM/MLLM that are aligned across languages or modalities and then (ii) uses them to build a language-agnostic or modality-agnostic classifier for detecting harmful prompts. Omniguard improves harmful prompt classification accuracy by 11.57\% over the strongest baseline in a multilingual setting, by 20.44\% for image-based prompts, and sets a new SOTA for audio-based prompts. By repurposing embeddings computed during generation, Omniguard is also very efficient ($\approx\!120 \times$ faster than the next fastest baseline). Code and data are available at: https://github.com/vsahil/OmniGuard.

As Artificial Intelligence (AI) systems, particularly those based on machine learning (ML), become integral to high-stakes applications, their probabilistic and opaque nature poses significant challenges to traditional verification and validation methods. These challenges are exacerbated in regulated sectors requiring tamper-proof, auditable evidence, as highlighted by apposite legal frameworks, e.g., the EU AI Act. Conversely, Zero-Knowledge Proofs (ZKPs) offer a cryptographic solution that enables provers to demonstrate, through verified computations, adherence to set requirements without revealing sensitive model details or data. Through a systematic survey of ZKP protocols, we identify five key properties (non-interactivity, transparent setup, standard representations, succinctness, and post-quantum security) critical for their application in AI validation and verification pipelines. Subsequently, we perform a follow-up systematic survey analyzing ZKP-enhanced ML applications across an adaptation of the Team Data Science Process (TDSP) model (Data & Preprocessing, Training & Offline Metrics, Inference, and Online Metrics), detailing verification objectives, ML models, and adopted protocols. Our findings indicate that current research on ZKP-Enhanced ML primarily focuses on inference verification, while the data preprocessing and training stages remain underexplored. Most notably, our analysis identifies a significant convergence within the research domain toward the development of a unified Zero-Knowledge Machine Learning Operations (ZKMLOps) framework. This emerging framework leverages ZKPs to provide robust cryptographic guarantees of correctness, integrity, and privacy, thereby promoting enhanced accountability, transparency, and compliance with Trustworthy AI principles.

People increasingly rely on AI-advice when making decisions. At times, such advice can promote selfish behavior. When individuals abide by selfishness-promoting AI advice, how are they perceived and punished? To study this question, we build on theories from social psychology and combine machine-behavior and behavioral economic approaches. In a pre-registered, financially-incentivized experiment, evaluators could punish real decision-makers who (i) received AI, human, or no advice. The advice (ii) encouraged selfish or prosocial behavior, and decision-makers (iii) behaved selfishly or, in a control condition, behaved prosocially. Evaluators further assigned responsibility to decision-makers and their advisors. Results revealed that (i) prosocial behavior was punished very little, whereas selfish behavior was punished much more. Focusing on selfish behavior, (ii) compared to receiving no advice, selfish behavior was penalized more harshly after prosocial advice and more leniently after selfish advice. Lastly, (iii) whereas selfish decision-makers were seen as more responsible when they followed AI compared to human advice, punishment between the two advice sources did not vary. Overall, behavior and advice content shape punishment, whereas the advice source does not.

Improving LLM First-Token Predictions in Multiple-Choice Question Answering via Output Prefilling

Silvia Cappelletti, Tobia Poppi, Samuele Poppi +5

Large Language Models (LLMs) are increasingly evaluated on multiple-choice question answering (MCQA) tasks using *first-token probability* (FTP), which selects the answer option whose initial token has the highest likelihood. While efficient, FTP can be fragile: models may assign high probability to unrelated tokens (*misalignment*) or use a valid token merely as part of a generic preamble rather than as a clear answer choice (*misinterpretation*), undermining the reliability of symbolic evaluation. We propose a simple solution: the *prefilling attack*, a structured natural-language prefix (e.g., "*The correct option is:*") prepended to the model output. Originally explored in AI safety, we repurpose prefilling to steer the model to respond with a clean, valid option, without modifying its parameters. Empirically, the FTP with prefilling strategy substantially improves accuracy, calibration, and output consistency across a broad set of LLMs and MCQA benchmarks. It outperforms standard FTP and often matches the performance of open-ended generation approaches that require full decoding and external classifiers, while being significantly more efficient. Our findings suggest that prefilling is a simple, robust, and low-cost method to enhance the reliability of FTP-based evaluation in multiple-choice settings.

Leaf diseases are harmful conditions that affect the health, appearance and productivity of plants, leading to significant plant loss and negatively impacting farmers' livelihoods. These diseases cause visible symptoms such as lesions, color changes, and texture variations, making it difficult for farmers to manage plant health, especially in large or remote farms where expert knowledge is limited. The main motivation of this study is to provide an efficient and accessible solution for identifying plant leaf diseases in Bangladesh, where agriculture plays a critical role in food security. The objective of our research is to classify 21 distinct leaf diseases across six plants using deep learning models, improving disease detection accuracy while reducing the need for expert involvement. Deep Learning (DL) techniques, including CNN and Transfer Learning (TL) models like VGG16, VGG19, MobileNetV2, InceptionV3, ResNet50V2 and Xception are used. VGG19 and Xception achieve the highest accuracies, with 98.90% and 98.66% respectively. Additionally, Explainable AI (XAI) techniques such as GradCAM, GradCAM++, LayerCAM, ScoreCAM and FasterScoreCAM are used to enhance transparency by highlighting the regions of the models focused on during disease classification. This transparency ensures that farmers can understand the model's predictions and take necessary action. This approach not only improves disease management but also supports farmers in making informed decisions, leading to better plant protection and increased agricultural productivity.

Innovations in digital intelligence are transforming robotic surgery with more informed decision-making. Real-time awareness of surgical instrument presence and actions (e.g., cutting tissue) is essential for such systems. Yet, despite decades of research, most machine learning models for this task are trained on small datasets and still struggle to generalize. Recently, vision-Language Models (VLMs) have brought transformative advances in reasoning across visual and textual modalities. Their unprecedented generalization capabilities suggest great potential for advancing intelligent robotic surgery. However, surgical VLMs remain under-explored, and existing models show limited performance, highlighting the need for benchmark studies to assess their capabilities and limitations and to inform future development. To this end, we benchmark the zero-shot performance of several advanced VLMs on two public robotic-assisted laparoscopic datasets for instrument and action classification. Beyond standard evaluation, we integrate explainable AI to visualize VLM attention and uncover causal explanations behind their predictions. This provides a previously underexplored perspective in this field for evaluating the reliability of model predictions. We also propose several explainability analysis-based metrics to complement standard evaluations. Our analysis reveals that surgical VLMs, despite domain-specific training, often rely on weak contextual cues rather than clinically relevant visual evidence, highlighting the need for stronger visual and reasoning supervision in surgical applications.

Rice is an essential staple food worldwide that is important in promoting international trade, economic growth, and nutrition. Asian countries such as China, India, Pakistan, Thailand, Vietnam, and Indonesia are notable for their significant contribution to the cultivation and utilization of rice. These nations are also known for cultivating different rice grains, including short and long grains. These sizes are further classified as basmati, jasmine, kainat saila, ipsala, arborio, etc., catering to diverse culinary preferences and cultural traditions. For both local and international trade, inspecting and maintaining the quality of rice grains to satisfy customers and preserve a country's reputation is necessary. Manual quality check and classification is quite a laborious and time-consuming process. It is also highly prone to mistakes. Therefore, an automatic solution must be proposed for the effective and efficient classification of different varieties of rice grains. This research paper presents an automatic framework based on a convolutional neural network (CNN) for classifying different varieties of rice grains. We evaluated the proposed model based on performance metrics such as accuracy, recall, precision, and F1-Score. The CNN model underwent rigorous training and validation, achieving a remarkable accuracy rate and a perfect area under each class's Receiver Operating Characteristic (ROC) curve. The confusion matrix analysis confirmed the model's effectiveness in distinguishing between the different rice varieties, indicating minimal misclassifications. Additionally, the integration of explainability techniques such as LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) provided valuable insights into the model's decision-making process, revealing how specific features of the rice grains influenced classification outcomes.