Large Language Models (LLMs) play a central role in modern artificial intelligence, yet their development has been primarily focused on English, resulting in limited support for other languages. We present PLLuM (Polish Large Language Model), the largest open-source family of foundation models tailored specifically for the Polish language. Developed by a consortium of major Polish research institutions, PLLuM addresses the need for high-quality, transparent, and culturally relevant language models beyond the English-centric commercial landscape. We describe the development process, including the construction of a new 140-billion-token Polish text corpus for pre-training, a 77k custom instructions dataset, and a 100k preference optimization dataset. A key component is a Responsible AI framework that incorporates strict data governance and a hybrid module for output correction and safety filtering. We detail the models' architecture, training procedures, and alignment techniques for both base and instruction-tuned variants, and demonstrate their utility in a downstream task within public administration. By releasing these models publicly, PLLuM aims to foster open research and strengthen sovereign AI technologies in Poland.

One major challenge for the adoption and acceptance of automated vehicles (AVs) is ensuring that they can make sound decisions in everyday situations that involve ethical tension. Much attention has focused on rare, high-stakes dilemmas such as trolley problems. Yet similar conflicts arise in routine driving when human considerations, such as legality, efficiency, and comfort, come into conflict. Current AV planning systems typically rely on rigid rules, which struggle to balance these competing considerations and often lead to behaviour that misaligns with human expectations. This paper introduces a reasons-based trajectory evaluation framework that operationalises the tracking condition of Meaningful Human Control (MHC). The framework represents human agents reasons (e.g., regulatory compliance) as quantifiable functions and evaluates how well candidate trajectories align with them. It assigns adjustable weights to agent priorities and includes a balance function to discourage excluding any agent. To demonstrate the approach, we use a real-world-inspired overtaking scenario, which highlights tensions between compliance, efficiency, and comfort. Our results show that different trajectories emerge as preferable depending on how agents reasons are weighted, and small shifts in priorities can lead to discrete changes in the selected action. This demonstrates that everyday ethical decisions in AV driving are highly sensitive to the weights assigned to the reasons of different human agents.

Autonomous driving systems remain brittle in rare, ambiguous, and out-of-distribution scenarios, where human driver succeed through contextual reasoning. Shared autonomy has emerged as a promising approach to mitigate such failures by incorporating human input when autonomy is uncertain. However, most existing methods restrict arbitration to low-level trajectories, which represent only geometric paths and therefore fail to preserve the underlying driving intent. We propose a unified shared autonomy framework that integrates human input and autonomous planners at a higher level of abstraction. Our method leverages Vision Language Models (VLMs) to infer driver intent from multi-modal cues -- such as driver actions and environmental context -- and to synthesize coherent strategies that mediate between human and autonomous control. We first study the framework in a mock-human setting, where it achieves perfect recall alongside high accuracy and precision. A human-subject survey further shows strong alignment, with participants agreeing with arbitration outcomes in 92% of cases. Finally, evaluation on the Bench2Drive benchmark demonstrates a substantial reduction in collision rate and improvement in overall performance compared to pure autonomy. Arbitration at the level of semantic, language-based representations emerges as a design principle for shared autonomy, enabling systems to exercise common-sense reasoning and maintain continuity with human intent.

The LLM-as-a-judge paradigm enables flexible, user-defined evaluation, but its effectiveness is often limited by the scarcity of diverse, representative data for refining criteria. We present a tool that integrates synthetic data generation into the LLM-as-a-judge workflow, empowering users to create tailored and challenging test cases with configurable domains, personas, lengths, and desired outcomes, including borderline cases. The tool also supports AI-assisted inline editing of existing test cases. To enhance transparency and interpretability, it reveals the prompts and explanations behind each generation. In a user study (N=24), 83% of participants preferred the tool over manually creating or selecting test cases, as it allowed them to rapidly generate diverse synthetic data without additional workload. The generated synthetic data proved as effective as hand-crafted data for both refining evaluation criteria and aligning with human preferences. These findings highlight synthetic data as a promising alternative, particularly in contexts where efficiency and scalability are critical.

We study a model of subscription-based platforms where users pay a fixed fee for unlimited access to content, and creators receive a share of the revenue. Existing approaches to detecting fraud predominantly rely on machine learning methods, engaging in an ongoing arms race with bad actors. We explore revenue division mechanisms that inherently disincentivize manipulation. We formalize three types of manipulation-resistance axioms and examine which existing rules satisfy these. We show that a mechanism widely used by streaming platforms, not only fails to prevent fraud, but also makes detecting manipulation computationally intractable. We also introduce a novel rule, ScaledUserProp, that satisfies all three manipulation-resistance axioms. Finally, experiments with both real-world and synthetic streaming data support ScaledUserProp as a fairer alternative compared to existing rules.

Time series classification is a task of paramount importance, as this kind of data often arises in safety-critical applications. However, it is typically tackled with black-box deep learning methods, making it hard for humans to understand the rationale behind their output. To take on this challenge, we propose a novel approach, STELLE (Signal Temporal logic Embedding for Logically-grounded Learning and Explanation), a neuro-symbolic framework that unifies classification and explanation through direct embedding of trajectories into a space of temporal logic concepts. By introducing a novel STL-inspired kernel that maps raw time series to their alignment with predefined STL formulae, our model jointly optimises accuracy and interpretability, as each prediction is accompanied by the most relevant logical concepts that characterise it. This yields (i) local explanations as human-readable STL conditions justifying individual predictions, and (ii) global explanations as class-characterising formulae. Experiments demonstrate that STELLE achieves competitive accuracy while providing logically faithful explanations, validated on diverse real-world benchmarks.

Machine unlearning aims to remove the influence of specific training data from a model without requiring full retraining. This capability is crucial for ensuring privacy, safety, and regulatory compliance. Therefore, verifying whether a model has truly forgotten target data is essential for maintaining reliability and trustworthiness. However, existing evaluation methods often assess forgetting at the level of individual inputs. This approach may overlook residual influence present in semantically similar examples. Such influence can compromise privacy and lead to indirect information leakage. We propose REMIND (Residual Memorization In Neighborhood Dynamics), a novel evaluation method aiming to detect the subtle remaining influence of unlearned data and classify whether the data has been effectively forgotten. REMIND analyzes the model's loss over small input variations and reveals patterns unnoticed by single-point evaluations. We show that unlearned data yield flatter, less steep loss landscapes, while retained or unrelated data exhibit sharper, more volatile patterns. REMIND requires only query-based access, outperforms existing methods under similar constraints, and demonstrates robustness across different models, datasets, and paraphrased inputs, making it practical for real-world deployment. By providing a more sensitive and interpretable measure of unlearning effectiveness, REMIND provides a reliable framework to assess unlearning in language models. As a result, REMIND offers a novel perspective on memorization and unlearning.

Large language models (LLMs) increasingly employ guardrails to enforce ethical, legal, and application-specific constraints on their outputs. While effective at mitigating harmful responses, these guardrails introduce a new class of vulnerabilities by exposing observable decision patterns. In this work, we present the first study of black-box LLM guardrail reverse-engineering attacks. We propose Guardrail Reverse-engineering Attack (GRA), a reinforcement learning-based framework that leverages genetic algorithm-driven data augmentation to approximate the decision-making policy of victim guardrails. By iteratively collecting input-output pairs, prioritizing divergence cases, and applying targeted mutations and crossovers, our method incrementally converges toward a high-fidelity surrogate of the victim guardrail. We evaluate GRA on three widely deployed commercial systems, namely ChatGPT, DeepSeek, and Qwen3, and demonstrate that it achieves an rule matching rate exceeding 0.92 while requiring less than $85 in API costs. These findings underscore the practical feasibility of guardrail extraction and highlight significant security risks for current LLM safety mechanisms. Our findings expose critical vulnerabilities in current guardrail designs and highlight the urgent need for more robust defense mechanisms in LLM deployment.

This study provides an empirical assessment of whether current large language models (LLMs) can pass the official qualifying examination for membership in Poland's National Appeal Chamber (Krajowa Izba Odwoławcza). The authors examine two related ideas: using LLM as actual exam candidates and applying the 'LLM-as-a-judge' approach, in which model-generated answers are automatically evaluated by other models. The paper describes the structure of the exam, which includes a multiple-choice knowledge test on public procurement law and a written judgment, and presents the hybrid information recovery and extraction pipeline built to support the models. Several LLMs (including GPT-4.1, Claude 4 Sonnet and Bielik-11B-v2.6) were tested in closed-book and various Retrieval-Augmented Generation settings. The results show that although the models achieved satisfactory scores in the knowledge test, none met the passing threshold in the practical written part, and the evaluations of the 'LLM-as-a-judge' often diverged from the judgments of the official examining committee. The authors highlight key limitations: susceptibility to hallucinations, incorrect citation of legal provisions, weaknesses in logical argumentation, and the need for close collaboration between legal experts and technical teams. The findings indicate that, despite rapid technological progress, current LLMs cannot yet replace human judges or independent examiners in Polish public procurement adjudication.

While the possibility of reaching human-like Artificial Intelligence (AI) remains controversial, the likelihood that the future will be characterized by a society with a growing presence of autonomous machines is high. Autonomous AI agents are already deployed and active across several industries and digital environments and alongside human-human and human-machine interactions, machine-machine interactions are poised to become increasingly prevalent. Given these developments, I argue that criminology must begin to address the implications of this transition for crime and social control. Drawing on Actor-Network Theory and Woolgar's decades-old call for a sociology of machines -- frameworks that acquire renewed relevance with the rise of generative AI agents -- I contend that criminologists should move beyond conceiving AI solely as a tool. Instead, AI agents should be recognized as entities with agency encompassing computational, social, and legal dimensions. Building on the literature on AI safety, I thus examine the risks associated with the rise of multi-agent AI systems, proposing a dual taxonomy to characterize the channels through which interactions among AI agents may generate deviant, unlawful, or criminal outcomes. I then advance and discuss four key questions that warrant theoretical and empirical attention: (1) Can we assume that machines will simply mimic humans? (2) Will crime theories developed for humans suffice to explain deviant or criminal behaviors emerging from interactions between autonomous AI agents? (3) What types of criminal behaviors will be affected first? (4) How might this unprecedented societal shift impact policing? These questions underscore the urgent need for criminologists to theoretically and empirically engage with the implications of multi-agent AI systems for the study of crime and play a more active role in debates on AI safety and governance.