Understanding how large language model (LLM) agents behave in strategic interactions is essential as these systems increasingly participate autonomously in economically and morally consequential decisions. We evaluate LLM preferences using canonical economic games, finding substantial deviations from human behavior. Models like GPT-4o show excessive cooperation and limited incentive sensitivity, while reasoning models, such as o3-mini, align more consistently with payoff-maximizing strategies. We propose a supervised fine-tuning pipeline that uses synthetic datasets derived from economic reasoning to align LLM agents with economic preferences, focusing on two stylized preference structures. In the first, utility depends only on individual payoffs (homo economicus), while utility also depends on a notion of Kantian universalizability in the second preference structure (homo moralis). We find that fine-tuning based on small datasets shifts LLM agent behavior toward the corresponding economic agent. We further assess the fine-tuned agents' behavior in two applications: Moral dilemmas involving autonomous vehicles and algorithmic pricing in competitive markets. These examples illustrate how different normative objectives embedded via realizations from structured preference structures can influence market and moral outcomes. This work contributes a replicable, cost-efficient, and economically grounded pipeline to align AI preferences using moral-economic principles.

Prevention of Future Deaths (PFD) reports, issued by coroners in England and Wales, flag systemic hazards that may lead to further loss of life. Analysis of these reports has previously been constrained by the manual effort required to identify and code relevant cases. In 2025, the Office for National Statistics (ONS) published a national thematic review of child-suicide PFD reports ($\leq$ 18 years), identifying 37 cases from January 2015 to November 2023 - a process based entirely on manual curation and coding. We evaluated whether a fully automated, open source "text-to-table" language-model pipeline (PFD Toolkit) could reproduce the ONS's identification and thematic analysis of child-suicide PFD reports, and assessed gains in efficiency and reliability. All 4,249 PFD reports published from July 2013 to November 2023 were processed via PFD Toolkit's large language model pipelines. Automated screening identified cases where the coroner attributed death to suicide in individuals aged 18 or younger, and eligible reports were coded for recipient category and 23 concern sub-themes, replicating the ONS coding frame. PFD Toolkit identified 72 child-suicide PFD reports - almost twice the ONS count. Three blinded clinicians adjudicated a stratified sample of 144 reports to validate the child-suicide screening. Against the post-consensus clinical annotations, the LLM-based workflow showed substantial to almost-perfect agreement (Cohen's $κ$ = 0.82, 95% CI: 0.66-0.98, raw agreement = 91%). The end-to-end script runtime was 8m 16s, transforming a process that previously took months into one that can be completed in minutes. This demonstrates that automated LLM analysis can reliably and efficiently replicate manual thematic reviews of coronial data, enabling scalable, reproducible, and timely insights for public health and safety. The PFD Toolkit is openly available for future research.

Smart contract comment generation has gained traction as a means to improve code comprehension and maintainability in blockchain systems. However, evaluating the quality of generated comments remains a challenge. Traditional metrics such as BLEU and ROUGE fail to capture domain-specific nuances, while human evaluation is costly and unscalable. In this paper, we present \texttt{evalSmarT}, a modular and extensible framework that leverages large language models (LLMs) as evaluators. The system supports over 400 evaluator configurations by combining approximately 40 LLMs with 10 prompting strategies. We demonstrate its application in benchmarking comment generation tools and selecting the most informative outputs. Our results show that prompt design significantly impacts alignment with human judgment, and that LLM-based evaluation offers a scalable and semantically rich alternative to existing methods.

Chain-of-Thought (CoT) prompting significantly enhances model reasoning, yet its internal mechanisms remain poorly understood. We analyze CoT's operational principles by reversely tracing information flow across decoding, projection, and activation phases. Our quantitative analysis suggests that CoT may serve as a decoding space pruner, leveraging answer templates to guide output generation, with higher template adherence strongly correlating with improved performance. Furthermore, we surprisingly find that CoT modulates neuron engagement in a task-dependent manner: reducing neuron activation in open-domain tasks, yet increasing it in closed-domain scenarios. These findings offer a novel mechanistic interpretability framework and critical insights for enabling targeted CoT interventions to design more efficient and robust prompts. We released our code and data at https://anonymous.4open.science/r/cot-D247.

The growing use of large language models (LLMs) has increased the need for automatic evaluation systems, particularly to address the challenge of information hallucination. Although existing faithfulness evaluation approaches have shown promise, they are predominantly English-focused and often require expensive human-labeled training data for fine-tuning specialized models. As LLMs see increased adoption in multilingual contexts, there is a need for accurate faithfulness evaluators that can operate across languages without extensive labeled data. This paper presents Self-Taught Evaluators for Multilingual Faithfulness, a framework that learns exclusively from synthetic multilingual summarization data while leveraging cross-lingual transfer learning. Through experiments comparing language-specific and mixed-language fine-tuning approaches, we demonstrate a consistent relationship between an LLM's general language capabilities and its performance in language-specific evaluation tasks. Our framework shows improvements over existing baselines, including state-of-the-art English evaluators and machine translation-based approaches.

While Multimodal Large Language Models (MLLMs) demonstrate impressive capabilities, their substantial computational and memory requirements pose significant barriers to practical deployment. Current parameter reduction techniques primarily involve training MLLMs from Small Language Models (SLMs), but these methods offer limited flexibility and remain computationally intensive. To address this gap, we propose to directly compress existing MLLMs through structural pruning combined with efficient recovery training. Specifically, we investigate two structural pruning paradigms--layerwise and widthwise pruning--applied to the language model backbone of MLLMs, alongside supervised finetuning and knowledge distillation. Additionally, we assess the feasibility of conducting recovery training with only a small fraction of the available data. Our results show that widthwise pruning generally maintains better performance in low-resource scenarios with limited computational resources or insufficient finetuning data. As for the recovery training, finetuning only the multimodal projector is sufficient at small compression levels (< 20%). Furthermore, a combination of supervised finetuning and hidden-state distillation yields optimal recovery across various pruning levels. Notably, effective recovery can be achieved with as little as 5% of the original training data, while retaining over 95% of the original performance. Through empirical study on two representative MLLMs, i.e., LLaVA-v1.5-7B and Bunny-v1.0-3B, this study offers actionable insights for practitioners aiming to compress MLLMs effectively without extensive computation resources or sufficient data.

Low-Rank Adaptation (LoRA) and its variants have delivered strong capability in Parameter-Efficient Transfer Learning (PETL) by minimizing trainable parameters and benefiting from reparameterization. However, their projection matrices remain unrestricted during training, causing high representation redundancy and diminishing the effectiveness of feature adaptation in the resulting subspaces. While existing methods mitigate this by manually adjusting the rank or implicitly applying channel-wise masks, they lack flexibility and generalize poorly across various datasets and architectures. Hence, we propose ReSoRA, a method that explicitly models redundancy between mapping subspaces and adaptively Regularizes Subspace redundancy of Low-Rank Adaptation. Specifically, it theoretically decomposes the low-rank submatrices into multiple equivalent subspaces and systematically applies de-redundancy constraints to the feature distributions across different projections. Extensive experiments validate that our proposed method consistently facilitates existing state-of-the-art PETL methods across various backbones and datasets in vision-language retrieval and standard visual classification benchmarks. Besides, as a training supervision, ReSoRA can be seamlessly integrated into existing approaches in a plug-and-play manner, with no additional inference costs. Code is publicly available at: https://github.com/Lucenova/ReSoRA.

The rapid spread of multilingual misinformation requires robust automated fact verification systems capable of handling fine-grained veracity assessments across diverse languages. While large language models have shown remarkable capabilities across many NLP tasks, their effectiveness for multilingual claim verification with nuanced classification schemes remains understudied. We conduct a comprehensive evaluation of five state-of-the-art language models on the X-Fact dataset, which spans 25 languages with seven distinct veracity categories. Our experiments compare small language models (encoder-based XLM-R and mT5) with recent decoder-only LLMs (Llama 3.1, Qwen 2.5, Mistral Nemo) using both prompting and fine-tuning approaches. Surprisingly, we find that XLM-R (270M parameters) substantially outperforms all tested LLMs (7-12B parameters), achieving 57.7% macro-F1 compared to the best LLM performance of 16.9%. This represents a 15.8% improvement over the previous state-of-the-art (41.9%), establishing new performance benchmarks for multilingual fact verification. Our analysis reveals problematic patterns in LLM behavior, including systematic difficulties in leveraging evidence and pronounced biases toward frequent categories in imbalanced data settings. These findings suggest that for fine-grained multilingual fact verification, smaller specialized models may be more effective than general-purpose large models, with important implications for practical deployment of fact-checking systems.

This paper proposes a method for generating machine-type-specific anomalies to evaluate the relative performance of unsupervised anomalous sound detection (UASD) systems across different machine types, even in the absence of real anomaly sound data. Conventional keyword-based data augmentation methods often produce unrealistic sounds due to their reliance on manually defined labels, limiting scalability as machine types and anomaly patterns diversify. Advanced audio generative models, such as MIMII-Gen, show promise but typically depend on anomalous training data, making them less effective when diverse anomalous examples are unavailable. To address these limitations, we propose a novel synthesis approach leveraging large language models (LLMs) to interpret textual descriptions of faults and automatically select audio transformation functions, converting normal machine sounds into diverse and plausible anomalous sounds. We validate this approach by evaluating a UASD system trained only on normal sounds from five machine types, using both real and synthetic anomaly data. Experimental results reveal consistent trends in relative detection difficulty across machine types between synthetic and real anomalies. This finding supports our hypothesis and highlights the effectiveness of the proposed LLM-based synthesis approach for relative evaluation of UASD systems.

Visual imitation learning (VIL) provides an efficient and intuitive strategy for robotic systems to acquire novel skills. Recent advancements in foundation models, particularly Vision Language Models (VLMs), have demonstrated remarkable capabilities in visual and linguistic reasoning for VIL tasks. Despite this progress, existing approaches primarily utilize these models for learning high-level plans from human demonstrations, relying on pre-defined motion primitives for executing physical interactions, which remains a major bottleneck for robotic systems. In this work, we present FMimic, a novel paradigm that harnesses foundation models to directly learn generalizable skills at even fine-grained action levels, using only a limited number of human videos. Extensive experiments demonstrate that our FMimic delivers strong performance with a single human video, and significantly outperforms all other methods with five videos. Furthermore, our method exhibits significant improvements of over 39% and 29% in RLBench multi-task experiments and real-world manipulation tasks, respectively, and exceeds baselines by more than 34% in high-precision tasks and 47% in long-horizon tasks.