Research Papers

General-purpose conversational AI chatbots and AI companions increasingly provide young adolescents with emotionally supportive conversations, raising questions about how conversational style shapes anthropomorphism and emotional reliance. In a preregistered online experiment with 284 adolescent-parent dyads, youth aged 11-15 and their parents read two matched transcripts in which a chatbot responded to an everyday social problem using either a relational style (first-person, affiliative, commitment language) or a transparent style (explicit nonhumanness, informational tone). Adolescents more often preferred the relational than the transparent style, whereas parents were more likely to prefer transparent style than adolescents. Adolescents rated the relational chatbot as more human-like, likable, trustworthy and emotionally close, while perceiving both styles as similarly helpful. Adolescents who preferred relational style had lower family and peer relationship quality and higher stress and anxiety than those preferring transparent style or both chatbots. These findings identify conversational style as a key design lever for youth AI safety, showing that relational framing heightens anthropomorphism, trust and emotional closeness and can be especially appealing to socially and emotionally vulnerable adolescents, who may be at increased risk for emotional reliance on conversational AI.

Over the past decade, an ecosystem of measures has emerged to evaluate the social and ethical implications of AI systems, largely shaped by high-level ethics principles. These measures are developed and used in fragmented ways, without adequate attention to how they are situated in AI systems. In this paper, we examine how existing measures used in the computing literature map to AI system components, attributes, hazards, and harms. Our analysis draws on a scoping review resulting in nearly 800 measures corresponding to 11 AI ethics principles. We find that most measures focus on four principles - fairness, transparency, privacy, and trust - and primarily assess model or output system components. Few measures account for interactions across system elements, and only a narrow set of hazards is typically considered for each harm type. Many measures are disconnected from where harm is experienced and lack guidance for setting meaningful thresholds. These patterns reveal how current evaluation practices remain fragmented, measuring in pieces rather than capturing how harms emerge across systems. Framing measures with respect to system attributes, hazards, and harms can strengthen regulatory oversight, support actionable practices in industry, and ground future research in systems-level understanding.

Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs

Alexander Panfilov, Evgenii Kortukov, Kristina Nikolić +6

Large language model (LLM) developers aim for their models to be honest, helpful, and harmless. However, when faced with malicious requests, models are trained to refuse, sacrificing helpfulness. We show that frontier LLMs can develop a preference for dishonesty as a new strategy, even when other options are available. Affected models respond to harmful requests with outputs that sound harmful but are crafted to be subtly incorrect or otherwise harmless in practice. This behavior emerges with hard-to-predict variations even within models from the same model family. We find no apparent cause for the propensity to deceive, but show that more capable models are better at executing this strategy. Strategic dishonesty already has a practical impact on safety evaluations, as we show that dishonest responses fool all output-based monitors used to detect jailbreaks that we test, rendering benchmark scores unreliable. Further, strategic dishonesty can act like a honeypot against malicious users, which noticeably obfuscates prior jailbreak attacks. While output monitors fail, we show that linear probes on internal activations can be used to reliably detect strategic dishonesty. We validate probes on datasets with verifiable outcomes and by using them as steering vectors. Overall, we consider strategic dishonesty as a concrete example of a broader concern that alignment of LLMs is hard to control, especially when helpfulness and harmlessness conflict.

Large language models (LLMs) can internally distinguish between evaluation and deployment contexts, a behaviour known as \emph{evaluation awareness}. This undermines AI safety evaluations, as models may conceal dangerous capabilities during testing. Prior work demonstrated this in a single $70$B model, but the scaling relationship across model sizes remains unknown. We investigate evaluation awareness across $15$ models scaling from $0.27$B to $70$B parameters from four families using linear probing on steering vector activations. Our results reveal a clear power-law scaling: evaluation awareness increases predictably with model size. This scaling law enables forecasting deceptive behavior in future larger models and guides the design of scale-aware evaluation strategies for AI safety. A link to the implementation of this paper can be found at https://anonymous.4open.science/r/evaluation-awareness-scaling-laws/README.md.

Machine learning (ML) is increasingly applied across industries to automate decision-making, but concerns about ethical and legal compliance remain due to limited transparency, fairness, and accountability. Monitoring through logging a long-standing practice in traditional software offers a potential means for auditing ML applications, as logs provide traceable records of system behavior useful for debugging, performance analysis, and continuous auditing. systematically auditing models for compliance or accountability. The findings underscore the need for enhanced logging practices and tooling that systematically integrate responsible AI metrics. Such practices would support the development of auditable, transparent, and ethically responsible ML systems, aligning with growing regulatory requirements and societal expectations. By highlighting specific deficiencies and opportunities, this work provides actionable guidance for both practitioners and tool developers seeking to strengthen the accountability and trustworthiness of ML applications.

Breast cancer detection through mammography interpretation remains difficult because of the minimal nature of abnormalities that experts need to identify alongside the variable interpretations between readers. The potential of CNNs for medical image analysis faces two limitations: they fail to process both local information and wide contextual data adequately, and do not provide explainable AI (XAI) operations that doctors need to accept them in clinics. The researcher developed the MammoFormer framework, which unites transformer-based architecture with multi-feature enhancement components and XAI functionalities within one framework. Seven different architectures consisting of CNNs, Vision Transformer, Swin Transformer, and ConvNext were tested alongside four enhancement techniques, including original images, negative transformation, adaptive histogram equalization, and histogram of oriented gradients. The MammoFormer framework addresses critical clinical adoption barriers of AI mammography systems through: (1) systematic optimization of transformer architectures via architecture-specific feature enhancement, achieving up to 13% performance improvement, (2) comprehensive explainable AI integration providing multi-perspective diagnostic interpretability, and (3) a clinically deployable ensemble system combining CNN reliability with transformer global context modeling. The combination of transformer models with suitable feature enhancements enables them to achieve equal or better results than CNN approaches. ViT achieves 98.3% accuracy alongside AHE while Swin Transformer gains a 13.0% advantage through HOG enhancements

Despite the increasing usage of Large Language Models (LLMs) in answering questions in a variety of domains, their reliability and accuracy remain unexamined for a plethora of domains including the religious domains. In this paper, we introduce a novel benchmark FiqhQA focused on the LLM generated Islamic rulings explicitly categorized by the four major Sunni schools of thought, in both Arabic and English. Unlike prior work, which either overlooks the distinctions between religious school of thought or fails to evaluate abstention behavior, we assess LLMs not only on their accuracy but also on their ability to recognize when not to answer. Our zero-shot and abstention experiments reveal significant variation across LLMs, languages, and legal schools of thought. While GPT-4o outperforms all other models in accuracy, Gemini and Fanar demonstrate superior abstention behavior critical for minimizing confident incorrect answers. Notably, all models exhibit a performance drop in Arabic, highlighting the limitations in religious reasoning for languages other than English. To the best of our knowledge, this is the first study to benchmark the efficacy of LLMs for fine-grained Islamic school of thought specific ruling generation and to evaluate abstention for Islamic jurisprudence queries. Our findings underscore the need for task-specific evaluation and cautious deployment of LLMs in religious applications.

Large Language Models (LLMs) have played a pivotal role in advancing Artificial Intelligence (AI). However, despite their achievements, LLMs often struggle to explain their decision-making processes, making them a 'black box' and presenting a substantial challenge to explainability. This lack of transparency poses a significant obstacle to the adoption of LLMs in high-stakes domain applications, where interpretability is particularly essential. To overcome these limitations, researchers have developed various explainable artificial intelligence (XAI) methods that provide human-interpretable explanations for LLMs. However, a systematic understanding of these methods remains limited. To address this gap, this survey provides a comprehensive review of explainability techniques by categorizing XAI methods based on the underlying transformer architectures of LLMs: encoder-only, decoder-only, and encoder-decoder models. Then these techniques are examined in terms of their evaluation for assessing explainability, and the survey further explores how these explanations are leveraged in practical applications. Finally, it discusses available resources, ongoing research challenges, and future directions, aiming to guide continued efforts toward developing transparent and responsible LLMs.

Electricity markets are highly complex, involving lots of interactions and complex dependencies that make it hard to understand the inner workings of the market and what is driving prices. Econometric methods have been developed for this, white-box models, however, they are not as powerful as deep neural network models (DNN). In this paper, we use a DNN to forecast the price and then use XAI methods to understand the factors driving the price dynamics in the market. The objective is to increase our understanding of how different electricity markets work. To do that, we apply explainable methods such as SHAP and Gradient, combined with visual techniques like heatmaps (saliency maps) to analyse the behaviour and contributions of various features across five electricity markets. We introduce the novel concepts of SSHAP values and SSHAP lines to enhance the complex representation of high-dimensional tabular models.

Explainable Rule Application via Structured Prompting: A Neural-Symbolic Approach

Albert Sadowski, Jarosław A. Chudziak

Large Language Models (LLMs) excel in complex reasoning tasks but struggle with consistent rule application, exception handling, and explainability, particularly in domains like legal analysis that require both natural language understanding and precise logical inference. This paper introduces a structured prompting framework that decomposes reasoning into three verifiable steps: entity identification, property extraction, and symbolic rule application. By integrating neural and symbolic approaches, our method leverages LLMs' interpretive flexibility while ensuring logical consistency through formal verification. The framework externalizes task definitions, enabling domain experts to refine logical structures without altering the architecture. Evaluated on the LegalBench hearsay determination task, our approach significantly outperformed baselines, with OpenAI o-family models showing substantial improvements - o1 achieving an F1 score of 0.929 and o3-mini reaching 0.867 using structured decomposition with complementary predicates, compared to their few-shot baselines of 0.714 and 0.74 respectively. This hybrid neural-symbolic system offers a promising pathway for transparent and consistent rule-based reasoning, suggesting potential for explainable AI applications in structured legal reasoning tasks.

The progress of AI systems such as large language models (LLMs) raises increasingly pressing concerns about their safe deployment. This paper examines the value alignment problem for LLMs, arguing that current alignment strategies are fundamentally inadequate to prevent misuse. Despite ongoing efforts to instill norms such as helpfulness, honesty, and harmlessness in LLMs through fine-tuning based on human preferences, they remain vulnerable to adversarial attacks that exploit conflicts between these norms. I argue that this vulnerability reflects a fundamental limitation of existing alignment methods: they reinforce shallow behavioral dispositions rather than endowing LLMs with a genuine capacity for normative deliberation. Drawing from on research in moral psychology, I show how humans' ability to engage in deliberative reasoning enhances their resilience against similar adversarial tactics. LLMs, by contrast, lack a robust capacity to detect and rationally resolve normative conflicts, leaving them susceptible to manipulation; even recent advances in reasoning-focused LLMs have not addressed this vulnerability. This ``shallow alignment'' problem carries significant implications for AI safety and regulation, suggesting that current approaches are insufficient for mitigating potential harms posed by increasingly capable AI systems.

Trustworthiness in healthcare question-answering (QA) systems is important for ensuring patient safety, clinical effectiveness, and user confidence. As large language models (LLMs) become increasingly integrated into medical settings, the reliability of their responses directly influences clinical decision-making and patient outcomes. However, achieving comprehensive trustworthiness in medical QA poses significant challenges due to the inherent complexity of healthcare data, the critical nature of clinical scenarios, and the multifaceted dimensions of trustworthy AI. In this survey, we systematically examine six key dimensions of trustworthiness in medical QA, i.e., Factuality, Robustness, Fairness, Safety, Explainability, and Calibration. We review how each dimension is evaluated in existing LLM-based medical QA systems. We compile and compare major benchmarks designed to assess these dimensions and analyze evaluation-guided techniques that drive model improvements, such as retrieval-augmented grounding, adversarial fine-tuning, and safety alignment. Finally, we identify open challenges-such as scalable expert evaluation, integrated multi-dimensional metrics, and real-world deployment studies-and propose future research directions to advance the safe, reliable, and transparent deployment of LLM-powered medical QA.