Research Papers

Alignment research focuses on making individual AI systems reliable. Human institutions achieve reliable collective behaviour differently: they mitigate the risk posed by misaligned individuals through organisational structure. Multi-agent AI systems should follow this institutional model using compartmentalisation and adversarial review to achieve reliable outcomes through architectural design rather than assuming individual alignment. We demonstrate this approach through the Perseverance Composition Engine, a multi-agent system for document composition. The Composer drafts text, the Corroborator verifies factual substantiation with full source access, and the Critic evaluates argumentative quality without access to sources: information asymmetry enforced by system architecture. This creates layered verification: the Corroborator detects unsupported claims, whilst the Critic independently assesses coherence and completeness. Observations from 474 composition tasks (discrete cycles of drafting, verification, and evaluation) exhibit patterns consistent with the institutional hypothesis. When assigned impossible tasks requiring fabricated content, this iteration enabled progression from attempted fabrication toward honest refusal with alternative proposals--behaviour neither instructed nor individually incentivised. These findings motivate controlled investigation of whether architectural enforcement produces reliable outcomes from unreliable components. This positions organisational theory as a productive framework for multi-agent AI safety. By implementing verification and evaluation as structural properties enforced through information compartmentalisation, institutional design offers a route to reliable collective behaviour from unreliable individual components.

Are AI Capabilities Increasing Exponentially? A Competing Hypothesis

Haosen Ge, Hamsa Bastani, Osbert Bastani

Rapidly increasing AI capabilities have substantial real-world consequences, ranging from AI safety concerns to labor market consequences. The Model Evaluation & Threat Research (METR) report argues that AI capabilities have exhibited exponential growth since 2019. In this note, we argue that the data does not support exponential growth, even in shorter-term horizons. Whereas the METR study claims that fitting sigmoid/logistic curves results in inflection points far in the future, we fit a sigmoid curve to their current data and find that the inflection point has already passed. In addition, we propose a more complex model that decomposes AI capabilities into base and reasoning capabilities, exhibiting individual rates of improvement. We prove that this model supports our hypothesis that AI capabilities will exhibit an inflection point in the near future. Our goal is not to establish a rigorous forecast of our own, but to highlight the fragility of existing forecasts of exponential growth.

Explainable Artificial Intelligence (XAI) has gained importance in interpreting model predictions. Among leading techniques for XAI, Local Interpretable Model-agnostic Explanations (LIME) is most frequently utilized as it notably helps people's understanding of complex models. However, LIME's analysis is constrained to a single image at a time. Besides, it lacks interaction mechanisms for observing the LIME's results and direct manipulations of factors affecting the results. To address these issues, we introduce an interactive visualization tool, LIMEVis, which improves the analysis workflow of LIME by enabling users to explore multiple LIME results simultaneously and modify them directly. With LIMEVis, we could conveniently identify common features in images that a model seems to mainly consider for category classification. Additionally, by interactively modifying the LIME results, we could determine which segments in an image influence the model's classification.

Online scams increasingly leverage fluent and context-aware social engineering strategies, creating growing demand for AI systems that explain why a message may be risky. However, explanations that cite detector-derived evidence may still semantically weaken or redirect the intended risk interpretation. We introduce VEXA: Verifying Semantic Explanation Alignment, a controlled testbed for studying the gap between lexical grounding and semantic risk alignment in AI-generated scam-risk explanations. VEXA generates ungrounded, risk-aligned, and risk-diluting explanations by independently controlling evidence grounding and semantic framing. Through LLM-as-a-judge and human evaluations, we show that explanations may continue to appear comparatively grounded even when their semantic interpretation weakens the detector's intended risk assessment. In human evaluation, risk-diluting XAI-grounded explanations retained comparatively elevated Perceived Evidence Grounding scores (3.66) despite lower Helpfulness (3.00) and Reasoning Support (3.14) scores. These findings provide controlled evidence of grounding illusion effects in AI-generated security explanations and suggest that trustworthy explanation evaluation must verify not only whether evidence is cited, but also how that evidence is interpreted.

Reliable Explanations or Random Noise? A Reliability Metric for XAI

Poushali Sengupta, Sabita Maharjan, Frank Eliassen +2

In recent years, explaining decisions made by complex machine learning models has become essential in high-stakes domains such as energy systems, healthcare, finance, and autonomous systems. However, the reliability of these explanations, namely, whether they remain stable and consistent under realistic, non-adversarial changes, remains largely unmeasured. Widely used methods such as SHAP and Integrated Gradients (IG) are well-motivated by axiomatic notions of attribution, yet their explanations can vary substantially even under system-level conditions, including small input perturbations, correlated representations, and minor model updates. Such variability undermines explanation reliability, as reliable explanations should remain consistent across equivalent input representations and small, performance-preserving model changes. We introduce the Explanation Reliability Index (ERI), a family of metrics that quantifies explanation stability under four reliability axioms: robustness to small input perturbations, consistency under feature redundancy, smoothness across model evolution, and resilience to mild distributional shifts. For each axiom, we derive formal guarantees, including Lipschitz-type bounds and temporal stability results. We further propose ERI-T, a dedicated measure of temporal reliability for sequential models, and introduce ERI-Bench, a benchmark designed to systematically stress-test explanation reliability across synthetic and real-world datasets. Experimental results reveal widespread reliability failures in popular explanation methods, showing that explanations can be unstable under realistic deployment conditions. By exposing and quantifying these instabilities, ERI enables principled assessment of explanation reliability and supports more trustworthy explainable AI (XAI) systems.

The increasing complexity and frequency of cyber-threats demand intrusion detection systems (IDS) that are not only accurate but also interpretable. This paper presented a novel IDS framework that integrated Explainable Artificial Intelligence (XAI) to enhance transparency in deep learning models. The framework was evaluated experimentally using the benchmark dataset NSL-KDD, demonstrating superior performance compared to traditional IDS and black-box deep learning models. The proposed approach combined Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) networks for capturing temporal dependencies in traffic sequences. Our deep learning results showed that both CNN and LSTM reached 0.99 for accuracy, whereas LSTM outperformed CNN at macro average precision, recall, and F-1 score. For weighted average precision, recall, and F-1 score, both models scored almost similarly. To ensure interpretability, the XAI model SHapley Additive exPlanations (SHAP) was incorporated, enabling security analysts to understand and validate model decisions. Some notable influential features were srv_serror_rate, dst_host_srv_serror_rate, and serror_rate for both models, as pointed out by SHAP. We also conducted a trust-focused expert survey based on IPIP6 and Big Five personality traits via an interactive UI to evaluate the system's reliability and usability. This work highlighted the potential of combining performance and transparency in cybersecurity solutions and recommends future enhancements through adaptive learning for real-time threat detection.

Large Language Models (LLMs) are increasingly deployed in critical applications requiring reliable reasoning, yet their internal reasoning processes remain difficult to evaluate systematically. Existing methods focus on final-answer correctness, providing limited insight into how reasoning unfolds across intermediate steps. We present EvalQReason, a framework that quantifies LLM reasoning quality through step-level probability distribution analysis without requiring human annotation. The framework introduces two complementary algorithms: Consecutive Step Divergence (CSD), which measures local coherence between adjacent reasoning steps, and Step-to-Final Convergence (SFC), which assesses global alignment with final answers. Each algorithm employs five statistical metrics to capture reasoning dynamics. Experiments across mathematical and medical datasets with open-source 7B-parameter models demonstrate that CSD-based features achieve strong predictive performance for correctness classification, with classical machine learning models reaching F1=0.78 and ROC-AUC=0.82, and sequential neural models substantially improving performance (F1=0.88, ROC-AUC=0.97). CSD consistently outperforms SFC, and sequential architectures outperform classical machine learning approaches. Critically, reasoning dynamics prove domain-specific: mathematical reasoning exhibits clear divergence-based discrimination patterns between correct and incorrect solutions, while medical reasoning shows minimal discriminative signals, revealing fundamental differences in how LLMs process different reasoning types. EvalQReason enables scalable, process-aware evaluation of reasoning reliability, establishing probability-based divergence analysis as a principled approach for trustworthy AI deployment.

Explainable artificial intelligence (XAI) aims to help uncover flaws in an AI model's internal representations. But do people draw the right conclusions from its explanations? Specifically, do they recognize an AI's inability to distinguish between relevant and irrelevant features? In the present study, a simulated AI classified images of railway trespassers as dangerous or not. To explain which features it has used, other images from the dataset were shown that activate the AI in a similar way. These concept images varied in three relevant features (i.e., a person's distance to the tracks, direction, and action) and in an irrelevant feature (i.e., scene background). When the AI uses a feature in its decision, this feature is retained in the concept images, otherwise the images randomize over it (e.g., same distance, varied backgrounds). Participants rated the AI more favorably when it retained relevant features. For the irrelevant feature, they did not mind in general, and sometimes even preferred it to be retained. This suggests that people may not recognize it when an AI model relies on irrelevant features to make its decisions.

Comparing Explanations is Not Enough, Explain the Change: New Standards are Needed to Explain Behavioral Shifts in Large Language Models

Martino Ciaperoni, Marzio Di Vece, Roberto Pellungrini +3

Large-scale foundation models exhibit behavioral shifts when subjected to interventions such as scaling, fine-tuning, reinforcement learning with human feedback, or in-context learning. Current explainability methods are structurally ill-suited to explain these shifts, because they either treat models as static objects, as traditional eXplainable AI (XAI) approaches do, or merely compare independent explanations across different checkpoints of a model. As a result, these approaches fail to explain the functional transition between two model instances in which a certain behavior has shifted following an intervention. This gap creates significant governance risks across jurisdictions including the EU AI Act, US state legislation, and Chinese AI regulations, which require documenting causal chains for substantial system modifications. This position paper argues that explaining behavioral shifts in large language models requires a principled approach that treats the shift itself as the primary object of explanation: namely, one that explains how and why an intervention transforms a reference model into an updated model with different behavior. To support this claim, we introduce Comparative XAI (XAI$_Δ$), a novel XAI paradigm aimed at explaining the difference between two model checkpoints where a behavior has shifted, together with a set of desiderata specifying what XAI$_Δ$ explainers and explanations must satisfy, including comparability, validity, actionability, and monitoring, with the goal of grounding model auditing in explicit, measurable requirements. Finally, we provide preliminary evidence suggesting the need for XAI$_Δ$ in practice through illustrative experiments, compiling the resulting findings into a transition report directly usable for governance and incident documentation.

As Generative AI becomes a key component in physics education, a significant ethical challenge has emerged: the tendency of students to anthropomorphize Large Language Models (LLMs), treating them as authoritative "oracles" that retrieve fixed facts from an internal database. However, LLMs operate fundamentally as probabilistic engines. This paper describes the design and implementation of a didactic activity, a reduced version of the "20 Questions" game, aimed at making this stochastic nature directly observable. Unlike a human player who fixes a target object at the start of the game, students discover that the model generates answers based solely on local coherence with the interaction history. By utilizing functionalities such as re-sampling and history rewinding, students act as experimenters, observing how identical interaction histories can yield diverging narrative paths. We discuss how mapping these behaviors to familiar physics concepts provides the epistemic scaffolding necessary to promote informed skepticism, framing the verification of AI outputs not merely as a compliance rule, but as a technical necessity derived from the system's probabilistic nature.

Artificial intelligence (AI) is increasingly used to support prognosis in Alzheimer's disease (AD), but adoption remains limited due to a lack of transparency and interpretability, particularly for long-term predictions where uncertainty is intrinsic and outcomes may not be known for years. We position uncertainty visualization as an explainable AI (XAI) technique and examine how it shapes trust, confidence, and reliance when users interpret AI-generated forecasts of future cognitive decline transitions. We conducted two studies, one with general participants (N=37) and one with experts in neuroimaging and neurology (N=10), to compare binary (present/absent) and continuous (saturation) uncertainty encodings. Continuous encodings improved perceived reliability and helped users recognize model limitations, while binary encodings increased momentary confidence, revealing expertise-dependent trade-offs in interpreting future predictions under high uncertainty. These findings surface key challenges in designing uncertainty representations for prognostic AI and culminate in a set of empirically grounded guidelines for creating trustworthy, user-appropriate clinical decision support tools.

PyGALAX is a Python package for geospatial analysis that integrates automated machine learning (AutoML) and explainable artificial intelligence (XAI) techniques to analyze spatial heterogeneity in both regression and classification tasks. It automatically selects and optimizes machine learning models for different geographic locations and contexts while maintaining interpretability through SHAP (SHapley Additive exPlanations) analysis. PyGALAX builds upon and improves the GALAX framework (Geospatial Analysis Leveraging AutoML and eXplainable AI), which has proven to outperform traditional geographically weighted regression (GWR) methods. Critical enhancements in PyGALAX from the original GALAX framework include automatic bandwidth selection and flexible kernel function selection, providing greater flexibility and robustness for spatial modeling across diverse datasets and research questions. PyGALAX not only inherits all the functionalities of the original GALAX framework but also packages them into an accessible, reproducible, and easily deployable Python toolkit while providing additional options for spatial modeling. It effectively addresses spatial non-stationarity and generates transparent insights into complex spatial relationships at both global and local scales, making advanced geospatial machine learning methods accessible to researchers and practitioners in geography, urban planning, environmental science, and related fields.