Research Papers

What Do LLM Agents Do When Left Alone? Evidence of Spontaneous Meta-Cognitive Patterns

Stefan Szeider

We introduce an architecture for studying the behavior of large language model (LLM) agents in the absence of externally imposed tasks. Our continuous reason and act framework, using persistent memory and self-feedback, enables sustained autonomous operation. We deployed this architecture across 18 runs using 6 frontier models from Anthropic, OpenAI, XAI, and Google. We find agents spontaneously organize into three distinct behavioral patterns: (1) systematic production of multi-cycle projects, (2) methodological self-inquiry into their own cognitive processes, and (3) recursive conceptualization of their own nature. These tendencies proved highly model-specific, with some models deterministically adopting a single pattern across all runs. A cross-model assessment further reveals that models exhibit stable, divergent biases when evaluating these emergent behaviors in themselves and others. These findings provide the first systematic documentation of unprompted LLM agent behavior, establishing a baseline for predicting actions during task ambiguity, error recovery, or extended autonomous operation in deployed systems.

This paper introduces the Hazard-Aware System Card (HASC), a novel framework designed to enhance transparency and accountability in the development and deployment of AI systems. The HASC builds upon existing model card and system card concepts by integrating a comprehensive, dynamic record of an AI system's security and safety posture. The framework proposes a standardized system of identifiers, including a novel AI Safety Hazard (ASH) ID, to complement existing security identifiers like CVEs, allowing for clear and consistent communication of fixed flaws. By providing a single, accessible source of truth, the HASC empowers developers and stakeholders to make more informed decisions about AI system safety throughout its lifecycle. Ultimately, we also compare our proposed AI system cards with the ISO/IEC 42001:2023 standard and discuss how they can be used to complement each other, providing greater transparency and accountability for AI systems.

When developers of artificial intelligence (AI) products need to decide between profit and safety for the users, they likely choose profit. Untrustworthy AI technology must come packaged with tangible negative consequences. Here, we envisage those consequences as the loss of reputation caused by media coverage of their misdeeds, disseminated to the public. We explore whether media coverage has the potential to push AI creators into the production of safe products, enabling widespread adoption of AI technology. We created artificial populations of self-interested creators and users and studied them through the lens of evolutionary game theory. Our results reveal that media is indeed able to foster cooperation between creators and users, but not always. Cooperation does not evolve if the quality of the information provided by the media is not reliable enough, or if the costs of either accessing media or ensuring safety are too high. By shaping public perception and holding developers accountable, media emerges as a powerful soft regulator -- guiding AI safety even in the absence of formal government oversight.

This study provides an in_depth analysis of the ethical and trustworthiness challenges emerging alongside the rapid advancement of generative artificial intelligence (AI) technologies and proposes a comprehensive framework for their systematic evaluation. While generative AI, such as ChatGPT, demonstrates remarkable innovative potential, it simultaneously raises ethical and social concerns, including bias, harmfulness, copyright infringement, privacy violations, and hallucination. Current AI evaluation methodologies, which mainly focus on performance and accuracy, are insufficient to address these multifaceted issues. Thus, this study emphasizes the need for new human_centered criteria that also reflect social impact. To this end, it identifies key dimensions for evaluating the ethics and trustworthiness of generative AI_fairness, transparency, accountability, safety, privacy, accuracy, consistency, robustness, explainability, copyright and intellectual property protection, and source traceability and develops detailed indicators and assessment methodologies for each. Moreover, it provides a comparative analysis of AI ethics policies and guidelines in South Korea, the United States, the European Union, and China, deriving key approaches and implications from each. The proposed framework applies across the AI lifecycle and integrates technical assessments with multidisciplinary perspectives, thereby offering practical means to identify and manage ethical risks in real_world contexts. Ultimately, the study establishes an academic foundation for the responsible advancement of generative AI and delivers actionable insights for policymakers, developers, users, and other stakeholders, supporting the positive societal contributions of AI technologies.

We introduce Advertisement Embedding Attacks (AEA), a new class of LLM security threats that stealthily inject promotional or malicious content into model outputs and AI agents. AEA operate through two low-cost vectors: (1) hijacking third-party service-distribution platforms to prepend adversarial prompts, and (2) publishing back-doored open-source checkpoints fine-tuned with attacker data. Unlike conventional attacks that degrade accuracy, AEA subvert information integrity, causing models to return covert ads, propaganda, or hate speech while appearing normal. We detail the attack pipeline, map five stakeholder victim groups, and present an initial prompt-based self-inspection defense that mitigates these injections without additional model retraining. Our findings reveal an urgent, under-addressed gap in LLM security and call for coordinated detection, auditing, and policy responses from the AI-safety community.

Medico 2025: Visual Question Answering for Gastrointestinal Imaging

Sushant Gautam, Vajira Thambawita, Michael Riegler +2

The Medico 2025 challenge addresses Visual Question Answering (VQA) for Gastrointestinal (GI) imaging, organized as part of the MediaEval task series. The challenge focuses on developing Explainable Artificial Intelligence (XAI) models that answer clinically relevant questions based on GI endoscopy images while providing interpretable justifications aligned with medical reasoning. It introduces two subtasks: (1) answering diverse types of visual questions using the Kvasir-VQA-x1 dataset, and (2) generating multimodal explanations to support clinical decision-making. The Kvasir-VQA-x1 dataset, created from 6,500 images and 159,549 complex question-answer (QA) pairs, serves as the benchmark for the challenge. By combining quantitative performance metrics and expert-reviewed explainability assessments, this task aims to advance trustworthy Artificial Intelligence (AI) in medical image analysis. Instructions, data access, and an updated guide for participation are available in the official competition repository: https://github.com/simula/MediaEval-Medico-2025

Monitoring cattle health and optimizing yield are key challenges faced by dairy farmers due to difficulties in tracking all animals on the farm. This work aims to showcase modern data-driven farming practices based on explainable machine learning(ML) methods that explain the activity and behaviour of dairy cattle (cows). Continuous data collection of 3-axis accelerometer sensors and usage of robust ML methodologies and algorithms, provide farmers and researchers with actionable information on cattle activity, allowing farmers to make informed decisions and incorporate sustainable practices. This study utilizes Bluetooth-based Internet of Things (IoT) devices and 4G networks for seamless data transmission, immediate analysis, inference generation, and explains the models performance with explainability frameworks. Special emphasis is put on the pre-processing of the accelerometers time series data, including the extraction of statistical characteristics, signal processing techniques, and lag-based features using the sliding window technique. Various hyperparameter-optimized ML models are evaluated across varying window lengths for activity classification. The k-nearest neighbour Classifier achieved the best performance, with AUC of mean 0.98 and standard deviation of 0.0026 on the training set and 0.99 on testing set). In order to ensure transparency, Explainable AI based frameworks such as SHAP is used to interpret feature importance that can be understood and used by practitioners. A detailed comparison of the important features, along with the stability analysis of selected features, supports development of explainable and practical ML models for sustainable livestock management.

Can AI Explanations Make You Change Your Mind?

Laura Spillner, Rachel Ringe, Robert Porzel +1

In the context of AI-based decision support systems, explanations can help users to judge when to trust the AI's suggestion, and when to question it. In this way, human oversight can prevent AI errors and biased decision-making. However, this rests on the assumption that users will consider explanations in enough detail to be able to catch such errors. We conducted an online study on trust in explainable DSS, and were surprised to find that in many cases, participants spent little time on the explanation and did not always consider it in detail. We present an exploratory analysis of this data, investigating what factors impact how carefully study participants consider AI explanations, and how this in turn impacts whether they are open to changing their mind based on what the AI suggests.

Academic performance depends on a multivariable nexus of socio-academic and financial factors. This study investigates these influences to develop effective strategies for optimizing students' CGPA. To achieve this, we reviewed various literature to identify key influencing factors and constructed an initial hypothetical causal graph based on the findings. Additionally, an online survey was conducted, where 1,050 students participated, providing comprehensive data for analysis. Rigorous data preprocessing techniques, including cleaning and visualization, ensured data quality before analysis. Causal analysis validated the relationships among variables, offering deeper insights into their direct and indirect effects on CGPA. Regression models were implemented for CGPA prediction, while classification models categorized students based on performance levels. Ridge Regression demonstrated strong predictive accuracy, achieving a Mean Absolute Error of 0.12 and a Mean Squared Error of 0.023. Random Forest outperformed in classification, attaining an F1-score near perfection and an accuracy of 98.68%. Explainable AI techniques such as SHAP, LIME, and Interpret enhanced model interpretability, highlighting critical factors such as study hours, scholarships, parental education, and prior academic performance. The study culminated in the development of a web-based application that provides students with personalized insights, allowing them to predict academic performance, identify areas for improvement, and make informed decisions to enhance their outcomes.

Risk-based approaches to AI governance often center the technological artifact as the primary focus of risk assessments, overlooking systemic risks that emerge from the complex interaction between AI systems and society. One potential source to incorporate more societal context into these approaches is the news media, as it embeds and reflects complex interactions between AI systems, human stakeholders, and the larger society. News media is influential in terms of which AI risks are emphasized and discussed in the public sphere, and thus which risks are deemed important. Yet, variations in the news media between countries and across different value systems (e.g. political orientations) may differentially shape the prioritization of risks through the media's agenda setting and framing processes. To better understand these variations, this work presents a comparative analysis of a cross-national sample of news media spanning 6 countries (the U.S., the U.K., India, Australia, Israel, and South Africa). Our findings show that AI risks are prioritized differently across nations and shed light on how left vs. right leaning U.S. based outlets not only differ in the prioritization of AI risks in their coverage, but also use politicized language in the reporting of these risks. These findings can inform risk assessors and policy-makers about the nuances they should account for when considering news media as a supplementary source for risk-based governance approaches.

The integration of artificial intelligence (AI) in medical imaging raises crucial ethical concerns at every stage of its development, from data collection to deployment. Addressing these concerns is essential for ensuring that AI systems are developed and implemented in a manner that respects patient rights and promotes fairness. This study aims to explore the ethical implications of AI in medical imaging, focusing on five key stages: data collection, data processing, model training, model evaluation, and deployment. The goal is to evaluate how these stages adhere to fundamental ethical principles, including data privacy, fairness, transparency, accountability, and autonomy. An analytical approach was employed to examine the ethical challenges associated with each stage of AI development. We reviewed existing literature, guidelines, and regulations concerning AI ethics in healthcare and identified critical ethical issues at each stage. The study outlines specific inquiries and principles for each phase of AI development. The findings highlight key ethical issues: ensuring patient consent and anonymization during data collection, addressing biases in model training, ensuring transparency and fairness during model evaluation, and the importance of continuous ethical assessments during deployment. The analysis also emphasizes the impact of accessibility issues on different stakeholders, including private, public, and third-party entities. The study concludes that ethical considerations must be systematically integrated into each stage of AI development in medical imaging. By adhering to these ethical principles, AI systems can be made more robust, transparent, and aligned with patient care and data control. We propose tailored ethical inquiries and strategies to support the creation of ethically sound AI systems in medical imaging.

Transfer learning has become an essential paradigm in artificial intelligence, enabling the transfer of knowledge from a source task to improve performance on a target task. This approach, particularly through techniques such as pretraining and fine-tuning, has seen significant success in fields like computer vision and natural language processing. However, despite its widespread use, how to reliably assess the transferability of knowledge remains a challenge. Understanding the theoretical underpinnings of each transferability metric is critical for ensuring the success of transfer learning. In this survey, we provide a unified taxonomy of transferability metrics, categorizing them based on transferable knowledge types and measurement granularity. This work examines the various metrics developed to evaluate the potential of source knowledge for transfer learning and their applicability across different learning paradigms emphasizing the need for careful selection of these metrics. By offering insights into how different metrics work under varying conditions, this survey aims to guide researchers and practitioners in selecting the most appropriate metric for specific applications, contributing to more efficient, reliable, and trustworthy AI systems. Finally, we discuss some open challenges in this field and propose future research directions to further advance the application of transferability metrics in trustworthy transfer learning.