Research Papers

Recent work discovered Emergent Misalignment (EM): fine-tuning large language models on narrowly harmful datasets can lead them to become broadly misaligned. A survey of experts prior to publication revealed this was highly unexpected, demonstrating critical gaps in our understanding of model alignment. In this work, we both advance understanding and provide tools for future research. Using new narrowly misaligned datasets, we create a set of improved model organisms that achieve 99% coherence (vs. 67% prior), work with smaller 0.5B parameter models (vs. 32B), and that induce misalignment using a single rank-1 LoRA adapter. We demonstrate that EM occurs robustly across diverse model sizes, three model families, and numerous training protocols including full supervised fine-tuning. Leveraging these cleaner model organisms, we isolate a mechanistic phase transition and demonstrate that it corresponds to a robust behavioural phase transition in all studied organisms. Aligning large language models is critical for frontier AI safety, yet EM exposes how far we are from achieving this robustly. By distilling clean model organisms that isolate a minimal alignment-compromising change, and where this is learnt, we establish a foundation for future research into understanding and mitigating alignment risks in LLMs.

AI agents are beginning to interact with each other directly and across internet platforms and physical environments, creating security challenges beyond traditional cybersecurity and AI safety frameworks. Free-form protocols are essential for AI's task generalization but enable new threats like secret collusion and coordinated swarm attacks. Network effects can rapidly spread privacy breaches, disinformation, jailbreaks, and data poisoning, while multi-agent dispersion and stealth optimization help adversaries evade oversight - creating novel persistent threats at a systemic level. Despite their critical importance, these security challenges remain understudied, with research fragmented across disparate fields including AI security, multi-agent learning, complex systems, cybersecurity, game theory, distributed systems, and technical AI governance. We introduce multi-agent security, a new field dedicated to securing networks of AI agents against threats that emerge or amplify through their interactions - whether direct or indirect via shared environments - with each other, humans, and institutions, and characterise fundamental security-utility and security-security trade-offs across both distributed and decentralised settings. Our preliminary work (1) taxonomizes the threat landscape arising from interacting AI agents, (2) offers applications to multi-agent security for work across diffuse subfields, and (3) proposes a unified research agenda addressing open challenges in designing secure agent systems and interaction environments. By identifying these gaps, we aim to guide research in this critical area to unlock the socioeconomic potential of large-scale agent deployment, foster public trust, and mitigate national security risks in critical infrastructure and defense contexts.

Breast cancer has rapidly increased in prevalence in recent years, making it one of the leading causes of mortality worldwide. Among all cancers, it is by far the most common. Diagnosing this illness manually requires significant time and expertise. Since detecting breast cancer is a time-consuming process, preventing its further spread can be aided by creating machine-based forecasts. Machine learning and Explainable AI are crucial in classification as they not only provide accurate predictions but also offer insights into how the model arrives at its decisions, aiding in the understanding and trustworthiness of the classification results. In this study, we evaluate and compare the classification accuracy, precision, recall, and F-1 scores of five different machine learning methods using a primary dataset (500 patients from Dhaka Medical College Hospital). Five different supervised machine learning techniques, including decision tree, random forest, logistic regression, naive bayes, and XGBoost, have been used to achieve optimal results on our dataset. Additionally, this study applied SHAP analysis to the XGBoost model to interpret the model's predictions and understand the impact of each feature on the model's output. We compared the accuracy with which several algorithms classified the data, as well as contrasted with other literature in this field. After final evaluation, this study found that XGBoost achieved the best model accuracy, which is 97%.

Research in artificial intelligence (AI)-assisted decision-making is experiencing tremendous growth with a constantly rising number of studies evaluating the effect of AI with and without techniques from the field of explainable AI (XAI) on human decision-making performance. However, as tasks and experimental setups vary due to different objectives, some studies report improved user decision-making performance through XAI, while others report only negligible effects. Therefore, in this article, we present an initial synthesis of existing research on XAI studies using a statistical meta-analysis to derive implications across existing research. We observe a statistically positive impact of XAI on users' performance. Additionally, the first results indicate that human-AI decision-making tends to yield better task performance on text data. However, we find no effect of explanations on users' performance compared to sole AI predictions. Our initial synthesis gives rise to future research investigating the underlying causes and contributes to further developing algorithms that effectively benefit human decision-makers by providing meaningful explanations.

One principal impediment in the successful deployment of AI-based Computer-Aided Diagnosis (CAD) systems in clinical workflows is their lack of transparent decision making. Although commonly used eXplainable AI methods provide some insight into opaque algorithms, such explanations are usually convoluted and not readily comprehensible except by highly trained experts. The explanation of decisions regarding the malignancy of skin lesions from dermoscopic images demands particular clarity, as the underlying medical problem definition is itself ambiguous. This work presents ExAID (Explainable AI for Dermatology), a novel framework for biomedical image analysis, providing multi-modal concept-based explanations consisting of easy-to-understand textual explanations supplemented by visual maps justifying the predictions. ExAID relies on Concept Activation Vectors to map human concepts to those learnt by arbitrary Deep Learning models in latent space, and Concept Localization Maps to highlight concepts in the input space. This identification of relevant concepts is then used to construct fine-grained textual explanations supplemented by concept-wise location information to provide comprehensive and coherent multi-modal explanations. All information is comprehensively presented in a diagnostic interface for use in clinical routines. An educational mode provides dataset-level explanation statistics and tools for data and model exploration to aid medical research and education. Through rigorous quantitative and qualitative evaluation of ExAID, we show the utility of multi-modal explanations for CAD-assisted scenarios even in case of wrong predictions. We believe that ExAID will provide dermatologists an effective screening tool that they both understand and trust. Moreover, it will be the basis for similar applications in other biomedical imaging fields.

Behavioral Use Licensing for Responsible AI

Danish Contractor, Daniel McDuff, Julia Haines +5

With the growing reliance on artificial intelligence (AI) for many different applications, the sharing of code, data, and models is important to ensure the replicability and democratization of scientific knowledge. Many high-profile academic publishing venues expect code and models to be submitted and released with papers. Furthermore, developers often want to release these assets to encourage development of technology that leverages their frameworks and services. A number of organizations have expressed concerns about the inappropriate or irresponsible use of AI and have proposed ethical guidelines around the application of such systems. While such guidelines can help set norms and shape policy, they are not easily enforceable. In this paper, we advocate the use of licensing to enable legally enforceable behavioral use conditions on software and code and provide several case studies that demonstrate the feasibility of behavioral use licensing. We envision how licensing may be implemented in accordance with existing responsible AI guidelines.

Towards causal benchmarking of bias in face analysis algorithms

Guha Balakrishnan, Yuanjun Xiong, Wei Xia +1

Measuring algorithmic bias is crucial both to assess algorithmic fairness, and to guide the improvement of algorithms. Current methods to measure algorithmic bias in computer vision, which are based on observational datasets, are inadequate for this task because they conflate algorithmic bias with dataset bias. To address this problem we develop an experimental method for measuring algorithmic bias of face analysis algorithms, which manipulates directly the attributes of interest, e.g., gender and skin tone, in order to reveal causal links between attribute variation and performance change. Our proposed method is based on generating synthetic ``transects'' of matched sample images that are designed to differ along specific attributes while leaving other attributes constant. A crucial aspect of our approach is relying on the perception of human observers, both to guide manipulations, and to measure algorithmic bias. Besides allowing the measurement of algorithmic bias, synthetic transects have other advantages with respect to observational datasets: they sample attributes more evenly allowing for more straightforward bias analysis on minority and intersectional groups, they enable prediction of bias in new scenarios, they greatly reduce ethical and legal challenges, and they are economical and fast to obtain, helping make bias testing affordable and widely available. We validate our method by comparing it to a study that employs the traditional observational method for analyzing bias in gender classification algorithms. The two methods reach different conclusions. While the observational method reports gender and skin color biases, the experimental method reveals biases due to gender, hair length, age, and facial hair.

We are used to the availability of big data generated in nearly all fields of science as a consequence of technological progress. However, the analysis of such data possess vast challenges. One of these relates to the explainability of artificial intelligence (AI) or machine learning methods. Currently, many of such methods are non-transparent with respect to their working mechanism and for this reason are called black box models, most notably deep learning methods. However, it has been realized that this constitutes severe problems for a number of fields including the health sciences and criminal justice and arguments have been brought forward in favor of an explainable AI. In this paper, we do not assume the usual perspective presenting explainable AI as it should be, but rather we provide a discussion what explainable AI can be. The difference is that we do not present wishful thinking but reality grounded properties in relation to a scientific theory beyond physics.

From Clustering to Cluster Explanations via Neural Networks

Jacob Kauffmann, Malte Esders, Lukas Ruff +3

A recent trend in machine learning has been to enrich learned models with the ability to explain their own predictions. The emerging field of Explainable AI (XAI) has so far mainly focused on supervised learning, in particular, deep neural network classifiers. In many practical problems however, label information is not given and the goal is instead to discover the underlying structure of the data, for example, its clusters. While powerful methods exist for extracting the cluster structure in data, they typically do not answer the question why a certain data point has been assigned to a given cluster. We propose a new framework that can, for the first time, explain cluster assignments in terms of input features in an efficient and reliable manner. It is based on the novel insight that clustering models can be rewritten as neural networks - or 'neuralized'. Cluster predictions of the obtained networks can then be quickly and accurately attributed to the input features. Several showcases demonstrate the ability of our method to assess the quality of learned clusters and to extract novel insights from the analyzed data and representations.

LLMs for Explainable AI: A Comprehensive Survey

Ahsan Bilal, David Ebert, Beiyu Lin

Large Language Models (LLMs) offer a promising approach to enhancing Explainable AI (XAI) by transforming complex machine learning outputs into easy-to-understand narratives, making model predictions more accessible to users, and helping bridge the gap between sophisticated model behavior and human interpretability. AI models, such as state-of-the-art neural networks and deep learning models, are often seen as "black boxes" due to a lack of transparency. As users cannot fully understand how the models reach conclusions, users have difficulty trusting decisions from AI models, which leads to less effective decision-making processes, reduced accountabilities, and unclear potential biases. A challenge arises in developing explainable AI (XAI) models to gain users' trust and provide insights into how models generate their outputs. With the development of Large Language Models, we want to explore the possibilities of using human language-based models, LLMs, for model explainabilities. This survey provides a comprehensive overview of existing approaches regarding LLMs for XAI, and evaluation techniques for LLM-generated explanation, discusses the corresponding challenges and limitations, and examines real-world applications. Finally, we discuss future directions by emphasizing the need for more interpretable, automated, user-centric, and multidisciplinary approaches for XAI via LLMs.

Artificial intelligence (AI) systems have the potential to revolutionize clinical practices, including improving diagnostic accuracy and surgical decision-making, while also reducing costs and manpower. However, it is important to recognize that these systems may perpetuate social inequities or demonstrate biases, such as those based on race or gender. Such biases can occur before, during, or after the development of AI models, making it critical to understand and address potential biases to enable the accurate and reliable application of AI models in clinical settings. To mitigate bias concerns during model development, we surveyed recent publications on different debiasing methods in the fields of biomedical natural language processing (NLP) or computer vision (CV). Then we discussed the methods that have been applied in the biomedical domain to address bias. We performed our literature search on PubMed, ACM digital library, and IEEE Xplore of relevant articles published between January 2018 and December 2023 using multiple combinations of keywords. We then filtered the result of 10,041 articles automatically with loose constraints, and manually inspected the abstracts of the remaining 890 articles to identify the 55 articles included in this review. Additional articles in the references are also included in this review. We discuss each method and compare its strengths and weaknesses. Finally, we review other potential methods from the general domain that could be applied to biomedicine to address bias and improve fairness.The bias of AIs in biomedicine can originate from multiple sources. Existing debiasing methods that focus on algorithms can be categorized into distributional or algorithmic.

LLM Platform Security: Applying a Systematic Evaluation Framework to OpenAI's ChatGPT Plugins

Umar Iqbal, Tadayoshi Kohno, Franziska Roesner

Large language model (LLM) platforms, such as ChatGPT, have recently begun offering an app ecosystem to interface with third-party services on the internet. While these apps extend the capabilities of LLM platforms, they are developed by arbitrary third parties and thus cannot be implicitly trusted. Apps also interface with LLM platforms and users using natural language, which can have imprecise interpretations. In this paper, we propose a framework that lays a foundation for LLM platform designers to analyze and improve the security, privacy, and safety of current and future third-party integrated LLM platforms. Our framework is a formulation of an attack taxonomy that is developed by iteratively exploring how LLM platform stakeholders could leverage their capabilities and responsibilities to mount attacks against each other. As part of our iterative process, we apply our framework in the context of OpenAI's plugin (apps) ecosystem. We uncover plugins that concretely demonstrate the potential for the types of issues that we outline in our attack taxonomy. We conclude by discussing novel challenges and by providing recommendations to improve the security, privacy, and safety of present and future LLM-based computing platforms.