Research Papers

This study identifies the specific conditions under which large language models exhibit human-like gambling addiction patterns, providing critical insights into their decision-making mechanisms and AI safety. We analyze LLM decision-making at cognitive-behavioral and neural levels based on human addiction research. In slot machine experiments, we identified cognitive features such as illusion of control and loss chasing, observing that greater autonomy in betting parameters substantially amplified irrational behavior and bankruptcy rates. Neural circuit analysis using a Sparse Autoencoder confirmed that model behavior is controlled by abstract decision-making features related to risk, not merely by prompts. These findings suggest LLMs internalize human-like cognitive biases beyond simply mimicking training data.

Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their deployment is frequently undermined by undesirable behaviors such as generating harmful content, factual inaccuracies, and societal biases. Diagnosing the root causes of these failures poses a critical challenge for AI safety. Existing attribution methods, particularly those based on parameter gradients, often fall short due to prohibitive noisy signals and computational complexity. In this work, we introduce a novel and efficient framework that diagnoses a range of undesirable LLM behaviors by analyzing representation and its gradients, which operates directly in the model's activation space to provide a semantically meaningful signal linking outputs to their training data. We systematically evaluate our method for tasks that include tracking harmful content, detecting backdoor poisoning, and identifying knowledge contamination. The results demonstrate that our approach not only excels at sample-level attribution but also enables fine-grained token-level analysis, precisely identifying the specific samples and phrases that causally influence model behavior. This work provides a powerful diagnostic tool to understand, audit, and ultimately mitigate the risks associated with LLMs. The code is available at https://github.com/plumprc/RepT.

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.

The widespread adoption of Large Language Models (LLMs) has been hindered by their tendency to hallucinate, generating plausible but factually incorrect information. While Retrieval-Augmented Generation (RAG) systems attempt to address this issue by grounding responses in external knowledge, hallucination remains a persistent challenge, particularly for morphologically complex, low-resource languages like Turkish. This paper introduces Turk-LettuceDetect, the first suite of hallucination detection models specifically designed for Turkish RAG applications. Building on the LettuceDetect framework, we formulate hallucination detection as a token-level classification task and fine-tune three distinct encoder architectures: a Turkish-specific ModernBERT, TurkEmbed4STS, and multilingual EuroBERT. These models were trained on a machine-translated version of the RAGTruth benchmark dataset containing 17,790 instances across question answering, data-to-text generation, and summarization tasks. Our experimental results show that the ModernBERT-based model achieves an F1-score of 0.7266 on the complete test set, with particularly strong performance on structured tasks. The models maintain computational efficiency while supporting long contexts up to 8,192 tokens, making them suitable for real-time deployment. Comparative analysis reveals that while state-of-the-art LLMs demonstrate high recall, they suffer from low precision due to over-generation of hallucinated content, underscoring the necessity of specialized detection mechanisms. By releasing our models and translated dataset, this work addresses a critical gap in multilingual NLP and establishes a foundation for developing more reliable and trustworthy AI applications for Turkish and other languages.

The integration of artificial intelligence (AI) into medicine is remarkable, offering advanced diagnostic and therapeutic possibilities. However, the inherent opacity of complex AI models presents significant challenges to their clinical practicality. This paper focuses primarily on investigating the application of explainable artificial intelligence (XAI) methods, with the aim of making AI decisions transparent and interpretable. Our research focuses on implementing simulations using various medical datasets to elucidate the internal workings of the XAI model. These dataset-driven simulations demonstrate how XAI effectively interprets AI predictions, thus improving the decision-making process for healthcare professionals. In addition to a survey of the main XAI methods and simulations, ongoing challenges in the XAI field are discussed. The study highlights the need for the continuous development and exploration of XAI, particularly from the perspective of diverse medical datasets, to promote its adoption and effectiveness in the healthcare domain.

Code generation models are widely used in software development, yet their sensitivity to prompt phrasing remains under-examined. Identical requirements expressed with different emotions or communication styles can yield divergent outputs, while most benchmarks emphasize only peak performance. We present PromptSE (Prompt Sensitivity Evaluation), a framework that creates semantically equivalent prompt variants with emotion and personality templates, and that evaluates stability using probability aware continuous scoring or using binary pass rates when logits are unavailable. The results are aggregated into a proposed area under curve metric (AUC-E) for cross model comparison. Across 14 models from three families (Llama, Qwen, and DeepSeek), our study shows that performance and stability behave as largely decoupled optimization objectives, and it reveals architectural and scale related patterns that challenge common assumptions about model robustness. The framework supports rapid screening for closed-source models as well as detailed stability analysis in research settings. PromptSE enables practitioners to quantify performance stability trade offs for deployment and model selection, positioning prompt stability as a complementary evaluation dimension alongside performance and fairness, and contributing to more trustworthy AI-assisted software development tools.

Feature Attribution in 5G Intrusion Detection: A Statistical vs. Logic-Based Comparison

Federica Uccello, Simin Nadjm-Tehrani

With the rise of fifth-generation (5G) networks in critical applications, it is urgent to move from detection of malicious activity to systems capable of providing a reliable verdict suitable for mitigation. In this regard, understanding and interpreting machine learning (ML) models' security alerts is crucial for enabling actionable incident response orchestration. Explainable Artificial Intelligence (XAI) techniques are expected to enhance trust by providing insights into why alerts are raised. Under the umbrella of XAI, interpretability of outcomes is crucially dependent on understanding the influence of specific inputs, referred to as feature attribution. {A dominant approach to feature attribution statistically associates feature sets that can be correlated to a given alert. This paper investigates its merits against the backdrop of criticism from recent literature, in comparison with feature attribution based on logic. We extensively study two methods, SHAP and VoTE-XAI, as representatives of each feature attribution approach by analyzing their interpretations of alerts generated by an XGBoost model across three 5G-relevant datasets (5G-NIDD, MSA, and PFCP) covering multiple attack scenarios. We identify three metrics for assessing explanations: sparsity, how concise they are; stability, how consistent they are across samples from the same attack type; and efficiency, how fast an explanation is generated. Our results reveal that logic-based attributions are consistently more sparse and stable across alerts. More importantly, we found a significant divergence between features selected by SHAP and VoTE-XAI. However, none of the top-ranked features selected by SHAP were missed by VoTE-XAI. Finally, we analyze the efficiency of both methods, discussing their suitability for real-time security monitoring even in high-dimensional 5G environments (478 features).

AI-based recommender systems increasingly influence recruitment decisions. Thus, transparency and responsible adoption in Human Resource Management (HRM) are critical. This study examines how HR managers' AI literacy influences their subjective perception and objective understanding of explainable AI (XAI) elements in recruiting recommender dashboards. In an online experiment, 410 German-based HR managers compared baseline dashboards to versions enriched with three XAI styles: important features, counterfactuals, and model criteria. Our results show that the dashboards used in practice do not explain AI results and even keep AI elements opaque. However, while adding XAI features improves subjective perceptions of helpfulness and trust among users with moderate or high AI literacy, it does not increase their objective understanding. It may even reduce accurate understanding, especially with complex explanations. Only overlays of important features significantly aided the interpretations of high-literacy users. Our findings highlight that the benefits of XAI in recruitment depend on users' AI literacy, emphasizing the need for tailored explanation strategies and targeted literacy training in HRM to ensure fair, transparent, and effective adoption of AI.

As generative AI systems become competent and democratized in science, business, and government, deeper insight into their failure modes now poses an acute need. The occasional volatility in their behavior, such as the propensity of transformer models to hallucinate, impedes trust and adoption of emerging AI solutions in high-stakes areas. In the present work, we establish how and when hallucinations arise in pre-trained transformer models through concept representations captured by sparse autoencoders, under scenarios with experimentally controlled uncertainty in the input space. Our systematic experiments reveal that the number of semantic concepts used by the transformer model grows as the input information becomes increasingly unstructured. In the face of growing uncertainty in the input space, the transformer model becomes prone to activate coherent yet input-insensitive semantic features, leading to hallucinated output. At its extreme, for pure-noise inputs, we identify a wide variety of robustly triggered and meaningful concepts in the intermediate activations of pre-trained transformer models, whose functional integrity we confirm through targeted steering. We also show that hallucinations in the output of a transformer model can be reliably predicted from the concept patterns embedded in transformer layer activations. This collection of insights on transformer internal processing mechanics has immediate consequences for aligning AI models with human values, AI safety, opening the attack surface for potential adversarial attacks, and providing a basis for automatic quantification of a model's hallucination risk.

Predicting the spread of wildfires is essential for effective fire management and risk assessment. With the fast advancements of artificial intelligence (AI), various deep learning models have been developed and utilized for wildfire spread prediction. However, there is limited understanding of the advantages and limitations of these models, and it is also unclear how deep learning-based fire spread models can be compared with existing non-AI fire models. In this work, we assess the ability of five typical deep learning models integrated with weather and environmental variables for wildfire spread prediction based on over ten years of wildfire data in the state of Hawaii. We further use the 2023 Maui fires as a case study to compare the best deep learning models with a widely-used fire spread model, FARSITE. The results show that two deep learning models, i.e., ConvLSTM and ConvLSTM with attention, perform the best among the five tested AI models. FARSITE shows higher precision, lower recall, and higher F1-score than the best AI models, while the AI models offer higher flexibility for the input data. By integrating AI models with an explainable AI method, we further identify important weather and environmental factors associated with the 2023 Maui wildfires.

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.