Research Papers

With origins in game theory, probabilistic values like Shapley values, Banzhaf values, and semi-values have emerged as a central tool in explainable AI. They are used for feature attribution, data attribution, data valuation, and more. Since exact computation can require exponential time, research has focused on efficient approximation methods using two techniques: Monte Carlo sampling and linear regression formulations. In this work, we present a new way of combining both of these techniques. Our approach is more flexible than prior algorithms, allowing for linear regression to be replaced with any function family whose probabilistic values can be computed efficiently. This allows us to harness the accuracy of tree-based models like XGBoost; with cross-fitting, the method also produces unbiased estimates. From experiments across eight datasets, we find that our methods give state-of-the-art performance for estimating probabilistic values. For Shapley values, the error of our methods can be $11\times$ lower than Permutation SHAP (the most popular Monte Carlo method), $6.5\times$ lower than Kernel SHAP (the most popular linear regression method), and $3.4\times$ lower than Leverage SHAP (the prior state-of-the-art Shapley value estimator). For more general probabilistic values, we can obtain error $390\times$ lower than the best estimator from prior work.

The honesty of large language models (LLMs) is a critical alignment challenge, especially as advanced systems with chain-of-thought (CoT) reasoning may strategically deceive humans. Unlike traditional honesty issues on LLMs, which could be possibly explained as some kind of hallucination, those models' explicit thought paths enable us to study strategic deception--goal-driven, intentional misinformation where reasoning contradicts outputs. Using representation engineering, we systematically induce, detect, and control such deception in CoT-enabled LLMs, extracting "deception vectors" via Linear Artificial Tomography (LAT) for 89% detection accuracy. Through activation steering, we achieve a 40% success rate in eliciting context-appropriate deception without explicit prompts, unveiling the specific honesty-related issue of reasoning models and providing tools for trustworthy AI alignment.

While the inconsistency of LLMs is not a novel topic, prior research has predominantly addressed two types of generative inconsistencies: i) Randomness Inconsistency: running the same LLM multiple trials, yielding varying responses; ii) Paraphrase Inconsistency: paraphrased prompts result in different responses from the same LLM. Randomness Inconsistency arises from the inherent randomness due to stochastic sampling in generative models, while Paraphrase Inconsistency is a consequence of the language modeling objectives, where paraphrased prompts alter the distribution of vocabulary logits. This research discovers Prompt-Reverse Inconsistency (PRIN), a new form of LLM self-inconsistency: given a question and a couple of LLM-generated answer candidates, the LLM often has conflicting responses when prompted "Which are correct answers?" and "Which are incorrect answers?". PRIN poses a big concern as it undermines the credibility of LLM-as-a-judge, and suggests a challenge for LLMs to adhere to basic logical rules. We conduct a series of experiments to investigate PRIN, examining the extent of PRIN across different LLMs, methods to mitigate it, potential applications, and its relationship with Randomness Inconsistency and Paraphrase Inconsistency. As the first study to explore PRIN, our findings offer valuable insights into the inner workings of LLMs and contribute to advancing trustworthy AI.

The growing reliance on Artificial Intelligence (AI) in critical domains such as healthcare demands robust mechanisms to ensure the trustworthiness of these systems, especially when faced with unexpected or anomalous inputs. This paper introduces the Open Medical Imaging Benchmarks for Out-Of-Distribution Detection (OpenMIBOOD), a comprehensive framework for evaluating out-of-distribution (OOD) detection methods specifically in medical imaging contexts. OpenMIBOOD includes three benchmarks from diverse medical domains, encompassing 14 datasets divided into covariate-shifted in-distribution, near-OOD, and far-OOD categories. We evaluate 24 post-hoc methods across these benchmarks, providing a standardized reference to advance the development and fair comparison of OOD detection methods. Results reveal that findings from broad-scale OOD benchmarks in natural image domains do not translate to medical applications, underscoring the critical need for such benchmarks in the medical field. By mitigating the risk of exposing AI models to inputs outside their training distribution, OpenMIBOOD aims to support the advancement of reliable and trustworthy AI systems in healthcare. The repository is available at https://github.com/remic-othr/OpenMIBOOD.

Explainable Artificial Intelligence (XAI) is essential for the transparency and clinical adoption of Clinical Decision Support Systems (CDSS). However, the real-world effectiveness of existing XAI methods remains limited and is inconsistently evaluated. This study conducts a systematic PRISMA-guided survey of 31 human-centered evaluations (HCE) of XAI applied to CDSS, classifying them by XAI methodology, evaluation design, and adoption barrier. Our findings reveal that most existing studies employ post-hoc, model-agnostic approaches such as SHAP and Grad-CAM, typically assessed through small-scale clinician studies. The results show that over 80% of the studies adopt post-hoc, model-agnostic approaches such as SHAP and Grad-CAM, and that clinician sample sizes remain below 25 participants. The findings indicate that explanations generally improve clinician trust and diagnostic confidence, but frequently increase cognitive load and exhibit misalignment with domain reasoning processes. To bridge these gaps, we propose a stakeholder-centric evaluation framework that integrates socio-technical principles and human-computer interaction to guide the future development of clinically viable and trustworthy XAI-based CDSS.

Examining the Expanding Role of Synthetic Data Throughout the AI Development Pipeline

Shivani Kapania, Stephanie Ballard, Alex Kessler +1

Alongside the growth of generative AI, we are witnessing a surge in the use of synthetic data across all stages of the AI development pipeline. It is now common practice for researchers and practitioners to use one large generative model (which we refer to as an auxiliary model) to generate synthetic data that is used to train or evaluate another, reconfiguring AI workflows and reshaping the very nature of data. While scholars have raised concerns over the risks of synthetic data, policy guidance and best practices for its responsible use have not kept up with these rapidly evolving industry trends, in part because we lack a clear picture of current practices and challenges. Our work aims to address this gap. Through 29 interviews with AI practitioners and responsible AI experts, we examine the expanding role of synthetic data in AI development. Our findings reveal how auxiliary models are now widely used across the AI development pipeline. Practitioners describe synthetic data as crucial for addressing data scarcity and providing a competitive edge, noting that evaluation of generative AI systems at scale would be infeasible without auxiliary models. However, they face challenges controlling the outputs of auxiliary models, generating data that accurately depict underrepresented groups, and scaling data validation practices that are based primarily on manual inspection. We detail general limitations of and ethical considerations for synthetic data and conclude with a proposal of concrete steps towards the development of best practices for its responsible use.

The joint implementation of federated learning (FL) and explainable artificial intelligence (XAI) could allow training models from distributed data and explaining their inner workings while preserving essential aspects of privacy. Toward establishing the benefits and tensions associated with their interplay, this scoping review maps the publications that jointly deal with FL and XAI, focusing on publications that reported an interplay between FL and model interpretability or post-hoc explanations. Out of the 37 studies meeting our criteria, only one explicitly and quantitatively analyzed the influence of FL on model explanations, revealing a significant research gap. The aggregation of interpretability metrics across FL nodes created generalized global insights at the expense of node-specific patterns being diluted. Several studies proposed FL algorithms incorporating explanation methods to safeguard the learning process against defaulting or malicious nodes. Studies using established FL libraries or following reporting guidelines are a minority. More quantitative research and structured, transparent practices are needed to fully understand their mutual impact and under which conditions it happens.

Federated Learning (FL) has become a key method for preserving data privacy in Internet of Things (IoT) environments, as it trains Machine Learning (ML) models locally while transmitting only model updates. Despite this design, FL remains susceptible to threats such as model inversion and membership inference attacks, which can reveal private training data. Differential Privacy (DP) techniques are often introduced to mitigate these risks, but simply injecting DP noise into black-box ML models can compromise accuracy, particularly in dynamic IoT contexts, where continuous, lifelong learning leads to excessive noise accumulation. To address this challenge, we propose Federated HyperDimensional computing with Privacy-preserving (FedHDPrivacy), an eXplainable Artificial Intelligence (XAI) framework that integrates neuro-symbolic computing and DP. Unlike conventional approaches, FedHDPrivacy actively monitors the cumulative noise across learning rounds and adds only the additional noise required to satisfy privacy constraints. In a real-world application for monitoring manufacturing machining processes, FedHDPrivacy maintains high performance while surpassing standard FL frameworks - Federated Averaging (FedAvg), Federated Proximal (FedProx), Federated Normalized Averaging (FedNova), and Federated Optimization (FedOpt) - by up to 37%. Looking ahead, FedHDPrivacy offers a promising avenue for further enhancements, such as incorporating multimodal data fusion.

While machine learning (ML) models have been able to achieve unprecedented accuracies across various prediction tasks in quantum chemistry, it is now apparent that accuracy on a test set alone is not a guarantee for robust chemical modeling such as stable molecular dynamics (MD). To go beyond accuracy, we use explainable artificial intelligence (XAI) techniques to develop a general analysis framework for atomic interactions and apply it to the SchNet and PaiNN neural network models. We compare these interactions with a set of fundamental chemical principles to understand how well the models have learned the underlying physicochemical concepts from the data. We focus on the strength of the interactions for different atomic species, how predictions for intensive and extensive quantum molecular properties are made, and analyze the decay and many-body nature of the interactions with interatomic distance. Models that deviate too far from known physical principles produce unstable MD trajectories, even when they have very high energy and force prediction accuracy. We also suggest further improvements to the ML architectures to better account for the polynomial decay of atomic interactions.

WalledEval is a comprehensive AI safety testing toolkit designed to evaluate large language models (LLMs). It accommodates a diverse range of models, including both open-weight and API-based ones, and features over 35 safety benchmarks covering areas such as multilingual safety, exaggerated safety, and prompt injections. The framework supports both LLM and judge benchmarking and incorporates custom mutators to test safety against various text-style mutations, such as future tense and paraphrasing. Additionally, WalledEval introduces WalledGuard, a new, small, and performant content moderation tool, and two datasets: SGXSTest and HIXSTest, which serve as benchmarks for assessing the exaggerated safety of LLMs and judges in cultural contexts. We make WalledEval publicly available at https://github.com/walledai/walledeval.

Ethical Framework for Responsible Foundational Models in Medical Imaging

Debesh Jha, Gorkem Durak, Abhijit Das +33

The emergence of foundational models represents a paradigm shift in medical imaging, offering extraordinary capabilities in disease detection, diagnosis, and treatment planning. These large-scale artificial intelligence systems, trained on extensive multimodal and multi-center datasets, demonstrate remarkable versatility across diverse medical applications. However, their integration into clinical practice presents complex ethical challenges that extend beyond technical performance metrics. This study examines the critical ethical considerations at the intersection of healthcare and artificial intelligence. Patient data privacy remains a fundamental concern, particularly given these models' requirement for extensive training data and their potential to inadvertently memorize sensitive information. Algorithmic bias poses a significant challenge in healthcare, as historical disparities in medical data collection may perpetuate or exacerbate existing healthcare inequities across demographic groups. The complexity of foundational models presents significant challenges regarding transparency and explainability in medical decision-making. We propose a comprehensive ethical framework that addresses these challenges while promoting responsible innovation. This framework emphasizes robust privacy safeguards, systematic bias detection and mitigation strategies, and mechanisms for maintaining meaningful human oversight. By establishing clear guidelines for development and deployment, we aim to harness the transformative potential of foundational models while preserving the fundamental principles of medical ethics and patient-centered care.

The black-box nature of deep learning models has raised concerns about their interpretability for successful deployment in real-world clinical applications. To address the concerns, eXplainable Artificial Intelligence (XAI) aims to provide clear and understandable explanations of the decision-making process. In the medical domain, concepts such as attributes of lesions or abnormalities serve as key evidence for deriving diagnostic results. Existing concept-based models mainly depend on concepts that appear independently and require fine-grained concept annotations such as bounding boxes. However, a medical image usually contains multiple concepts, and the fine-grained concept annotations are difficult to acquire. In this paper, we aim to interpret representations in deep neural networks by aligning the axes of the latent space with known concepts of interest. We propose a novel Concept-Attention Whitening (CAW) framework for interpretable skin lesion diagnosis. CAW is comprised of a disease diagnosis branch and a concept alignment branch. In the former branch, we train a convolutional neural network (CNN) with an inserted CAW layer to perform skin lesion diagnosis. The CAW layer decorrelates features and aligns image features to conceptual meanings via an orthogonal matrix. In the latter branch, the orthogonal matrix is calculated under the guidance of the concept attention mask. We particularly introduce a weakly-supervised concept mask generator that only leverages coarse concept labels for filtering local regions that are relevant to certain concepts, improving the optimization of the orthogonal matrix. Extensive experiments on two public skin lesion diagnosis datasets demonstrated that CAW not only enhanced interpretability but also maintained a state-of-the-art diagnostic performance.