Due to their ability to process long and complex contexts, LLMs can offer key benefits to the Legal domain, but their adoption has been hindered by their tendency to generate unfaithful, ungrounded, or hallucinatory outputs. While Retrieval-Augmented Generation offers a promising solution by grounding generations in external knowledge, it offers no guarantee that the provided context will be effectively integrated. To address this, context-aware decoding strategies have been proposed to amplify the influence of relevant context, but they usually do not explicitly enforce faithfulness to the context. In this work, we introduce Confidence-guided Copy-based Decoding for Legal Text Generation (CoCoLex)-a decoding strategy that dynamically interpolates the model produced vocabulary distribution with a distribution derived based on copying from the context. CoCoLex encourages direct copying based on the model's confidence, ensuring greater fidelity to the source. Experimental results on five legal benchmarks demonstrate that CoCoLex outperforms existing context-aware decoding methods, particularly in long-form generation tasks.

As automated vehicles (AVs) increasingly integrate into mixed-traffic environments, evaluating their interaction with human-driven vehicles (HDVs) becomes critical. In most research focused on developing new AV control algorithms (controllers), the performance of these algorithms is assessed solely based on performance metrics such as collision avoidance or lane-keeping efficiency, while largely overlooking the human-centred dimensions of interaction with HDVs. This paper proposes a structured evaluation framework that addresses this gap by incorporating metrics grounded in the human-robot interaction literature. The framework spans four key domains: a) interaction effect, b) interaction perception, c) interaction effort, and d) interaction ability. These domains capture both the performance of the AV and its impact on human drivers around it. To demonstrate the utility of the framework, we apply it to a case study evaluating how a state-of-the-art AV controller interacts with human drivers in a merging scenario in a driving simulator. Measuring HDV-HDV interactions as a baseline, this study included one representative metric per domain: a) perceived safety, b) subjective ratings, specifically how participants perceived the other vehicle's driving behaviour (e.g., aggressiveness or predictability) , c) driver workload, and d) merging success. The results showed that incorporating metrics covering all four domains in the evaluation of AV controllers can illuminate critical differences in driver experience when interacting with AVs. This highlights the need for a more comprehensive evaluation approach. Our framework offers researchers, developers, and policymakers a systematic method for assessing AV behaviour beyond technical performance, fostering the development of AVs that are not only functionally capable but also understandable, acceptable, and safe from a human perspective.

Reasoning with minimal models has always been at the core of many knowledge representation techniques, but we still have only a limited understanding of this problem in Description Logics (DLs). Minimization of some selected predicates, letting the remaining predicates vary or be fixed, as proposed in circumscription, has been explored and exhibits high complexity. The case of `pure' minimal models, where the extension of all predicates must be minimal, has remained largely uncharted. We address this problem in popular DLs and obtain surprisingly negative results: concept satisfiability in minimal models is undecidable already for $\mathcal{EL}$. This undecidability also extends to a very restricted fragment of tuple-generating dependencies. To regain decidability, we impose acyclicity conditions on the TBox that bring the worst-case complexity below double exponential time and allow us to establish a connection with the recently studied pointwise circumscription; we also derive results in data complexity. We conclude with a brief excursion to the DL-Lite family, where a positive result was known for DL-Lite$_{\text{core}}$, but our investigation establishes ExpSpace-hardness already for its extension DL-Lite$_{\text{horn}}$.

Timely and effective incident response is key to managing the growing frequency of cyberattacks. However, identifying the right response actions for complex systems is a major technical challenge. A promising approach to mitigate this challenge is to use the security knowledge embedded in large language models (LLMs) to assist security operators during incident handling. Recent research has demonstrated the potential of this approach, but current methods are mainly based on prompt engineering of frontier LLMs, which is costly and prone to hallucinations. We address these limitations by presenting a novel way to use an LLM for incident response planning with reduced hallucination. Our method includes three steps: fine-tuning, information retrieval, and lookahead planning. We prove that our method generates response plans with a bounded probability of hallucination and that this probability can be made arbitrarily small at the expense of increased planning time under certain assumptions. Moreover, we show that our method is lightweight and can run on commodity hardware. We evaluate our method on logs from incidents reported in the literature. The experimental results show that our method a) achieves up to 22% shorter recovery times than frontier LLMs and b) generalizes to a broad range of incident types and response actions.

This paper introduces a comprehensive benchmark for evaluating how Large Language Models (LLMs) respond to linguistic shibboleths: subtle linguistic markers that can inadvertently reveal demographic attributes such as gender, social class, or regional background. Through carefully constructed interview simulations using 100 validated question-response pairs, we demonstrate how LLMs systematically penalize certain linguistic patterns, particularly hedging language, despite equivalent content quality. Our benchmark generates controlled linguistic variations that isolate specific phenomena while maintaining semantic equivalence, which enables the precise measurement of demographic bias in automated evaluation systems. We validate our approach along multiple linguistic dimensions, showing that hedged responses receive 25.6% lower ratings on average, and demonstrate the benchmark's effectiveness in identifying model-specific biases. This work establishes a foundational framework for detecting and measuring linguistic discrimination in AI systems, with broad applications to fairness in automated decision-making contexts.

The internet is rife with unattributed, deliberately misleading, or otherwise untrustworthy content. Though large language models (LLMs) are often tasked with autonomous web browsing, the extent to which they have learned the simple heuristics human researchers use to navigate this noisy environment is not currently known. In this paper, we introduce the Synthetic Media Literacy Test (SMeL Test), a minimal benchmark that tests the ability of language models to actively filter out untrustworthy information in context. We benchmark a variety of commonly used instruction-tuned LLMs, including reasoning models, and find that no model consistently succeeds; while reasoning in particular is associated with higher scores, even the best API model we test hallucinates up to 70% of the time. Remarkably, larger and more capable models do not necessarily outperform their smaller counterparts. We hope our work sheds more light on this important form of hallucination and guides the development of new methods to combat it.

The ongoing evolution of AI paradigms has propelled AI research into the agentic AI stage. Consequently, the focus of research has shifted from single agents and simple applications towards multi-agent autonomous decision-making and task collaboration in complex environments. As Large Language Models (LLMs) advance, their applications become more diverse and complex, leading to increasing situational and systemic risks. This has brought significant attention to value alignment for agentic AI systems, which aims to ensure that an agent's goals, preferences, and behaviors align with human values and societal norms. Addressing socio-governance demands through a Multi-level Value framework, this study comprehensively reviews value alignment in LLM-based multi-agent systems as the representative archetype of agentic AI systems. Our survey systematically examines three interconnected dimensions: First, value principles are structured via a top-down hierarchy across macro, meso, and micro levels. Second, application scenarios are categorized along a general-to-specific continuum explicitly mirroring these value tiers. Third, value alignment methods and evaluation are mapped to this tiered framework through systematic examination of benchmarking datasets and relevant methodologies. Additionally, we delve into value coordination among multiple agents within agentic AI systems. Finally, we propose several potential research directions in this field.

Large Language Models (LLMs) are trained with safety alignment to prevent generating malicious content. Although some attacks have highlighted vulnerabilities in these safety-aligned LLMs, they typically have limitations, such as necessitating access to the model weights or the generation process. Since proprietary models through API-calling do not grant users such permissions, these attacks find it challenging to compromise them. In this paper, we propose Jailbreaking Using LLM Introspection (JULI), which jailbreaks LLMs by manipulating the token log probabilities, using a tiny plug-in block, BiasNet. JULI relies solely on the knowledge of the target LLM's predicted token log probabilities. It can effectively jailbreak API-calling LLMs under a black-box setting and knowing only top-$5$ token log probabilities. Our approach demonstrates superior effectiveness, outperforming existing state-of-the-art (SOTA) approaches across multiple metrics.

Stability guarantees have emerged as a principled way to evaluate feature attributions, but existing certification methods rely on heavily smoothed classifiers and often produce conservative guarantees. To address these limitations, we introduce soft stability and propose a simple, model-agnostic, sample-efficient stability certification algorithm (SCA) that yields non-trivial and interpretable guarantees for any attribution method. Moreover, we show that mild smoothing achieves a more favorable trade-off between accuracy and stability, avoiding the aggressive compromises made in prior certification methods. To explain this behavior, we use Boolean function analysis to derive a novel characterization of stability under smoothing. We evaluate SCA on vision and language tasks and demonstrate the effectiveness of soft stability in measuring the robustness of explanation methods.

Speculative Decoding (SD) is a widely used approach to accelerate the inference of large language models (LLMs) without reducing generation quality. It operates by first using a compact model to draft multiple tokens efficiently, followed by parallel verification using the target LLM. This approach leads to faster inference compared to auto-regressive decoding. While there are multiple approaches to create a draft model, one promising approach is to use early-exit methods. These methods draft candidate tokens by using a subset of layers of the primary model and applying the remaining layers for verification, allowing a single model to handle both drafting and verification. While this technique reduces memory usage and computational cost, its performance relies on the choice of the exit layer for drafting and the number of tokens drafted (speculation length) in each SD round. Prior works use hyperparameter exploration to statically select these values. However, our evaluations show that these hyperparameter values are task-specific, and even within a task they are dependent on the current sequence context. We introduce DEL (Dynamic Exit Layer), a plug-and-play method that adaptively selects the exit layer and speculation length during inference. DEL dynamically tracks the token acceptance rate if the tokens are drafted at each layer of an LLM and uses that knowledge to heuristically select the optimal exit layer and speculation length. Our experiments across a broad range of models and downstream tasks show that DEL achieves overall speedups of $2.16\times$$\sim$$2.62\times$ over vanilla auto-regressive decoding and improves upon state-of-the-art SD methods, which peak at $2.43\times$, by up to $0.19\times$. The code is available at https://github.com/hoenza/DEL.