Safety alignment in Large Language Models (LLMs) often involves mediating internal representations to refuse harmful requests. Recent research has demonstrated that these safety mechanisms can be bypassed by ablating or removing specific representational directions within the model. In this paper, we propose the opposite approach: Rank-One Safety Injection (ROSI), a white-box method that amplifies a model's safety alignment by permanently steering its activations toward the refusal-mediating subspace. ROSI operates as a simple, fine-tuning-free rank-one weight modification applied to all residual stream write matrices. The required safety direction can be computed from a small set of harmful and harmless instruction pairs. We show that ROSI consistently increases safety refusal rates - as evaluated by Llama Guard 3 - while preserving the utility of the model on standard benchmarks such as MMLU, HellaSwag, and Arc. Furthermore, we show that ROSI can also re-align 'uncensored' models by amplifying their own latent safety directions, demonstrating its utility as an effective last-mile safety procedure. Our results suggest that targeted, interpretable weight steering is a cheap and potent mechanism to improve LLM safety, complementing more resource-intensive fine-tuning paradigms.

Recent advances in large language models (LLMs) and medical LLMs (Med-LLMs) have demonstrated strong performance on general medical benchmarks. However, their capabilities in specialized medical fields, such as dentistry which require deeper domain-specific knowledge, remain underexplored due to the lack of targeted evaluation resources. In this paper, we introduce DentalBench, the first comprehensive bilingual benchmark designed to evaluate and advance LLMs in the dental domain. DentalBench consists of two main components: DentalQA, an English-Chinese question-answering (QA) benchmark with 36,597 questions spanning 4 tasks and 16 dental subfields; and DentalCorpus, a large-scale, high-quality corpus with 337.35 million tokens curated for dental domain adaptation, supporting both supervised fine-tuning (SFT) and retrieval-augmented generation (RAG). We evaluate 14 LLMs, covering proprietary, open-source, and medical-specific models, and reveal significant performance gaps across task types and languages. Further experiments with Qwen-2.5-3B demonstrate that domain adaptation substantially improves model performance, particularly on knowledge-intensive and terminology-focused tasks, and highlight the importance of domain-specific benchmarks for developing trustworthy and effective LLMs tailored to healthcare applications.

Psychometric tests, traditionally used to assess humans, are now being applied to Large Language Models (LLMs) to evaluate their behavioral traits. However, existing studies follow a context-free approach, answering each question in isolation to avoid contextual influence. We term this the Disney World test, an artificial setting that ignores real-world applications, where conversational history shapes responses. To bridge this gap, we propose the first Context-Aware Personality Evaluation (CAPE) framework for LLMs, incorporating prior conversational interactions. To thoroughly analyze the influence of context, we introduce novel metrics to quantify the consistency of LLM responses, a fundamental trait in human behavior. Our exhaustive experiments on 7 LLMs reveal that conversational history enhances response consistency via in-context learning but also induces personality shifts, with GPT-3.5-Turbo and GPT-4-Turbo exhibiting extreme deviations. While GPT models are robust to question ordering, Gemini-1.5-Flash and Llama-8B display significant sensitivity. Moreover, GPT models response stem from their intrinsic personality traits as well as prior interactions, whereas Gemini-1.5-Flash and Llama--8B heavily depend on prior interactions. Finally, applying our framework to Role Playing Agents (RPAs) shows context-dependent personality shifts improve response consistency and better align with human judgments. Our code and datasets are publicly available at: https://github.com/jivnesh/CAPE

--The impressive performance of transformer models has sparked the deployment of intelligent applications on resource-constrained edge devices. However, ensuring high-quality service for real-time edge systems is a significant challenge due to the considerable computational demands and resource requirements of these models. Existing strategies typically either offload transformer computations to other devices or directly deploy compressed models on individual edge devices. T o tackle these challenges, we propose a collaborative inference system for general transformer models, termed CoFormer . The central idea behind CoFormer is to exploit the divisibility and integrability of transformer . An off-the-shelf large transformer can be decomposed into multiple smaller models for distributed inference, and their intermediate results are aggregated to generate the final output. We formulate an optimization problem to minimize both inference latency and accuracy degradation under heterogeneous hardware constraints. DeBo algorithm is proposed to first solve the optimization problem to derive the decomposition policy, and then progressively calibrate decomposed models to restore performance. We demonstrate the capability to support a wide range of transformer models on heterogeneous edge devices, achieving up to 3.1 inference speedup with large transformer models. Notably, CoFormer enables the efficient inference of GPT2-XL with 1.6 billion parameters on edge devices, reducing memory requirements by 76.3%. CoFormer can also reduce energy consumption by approximately 40% while maintaining satisfactory inference performance. Guanyu Xu, Zhiwei Hao and Han Hu are with the School of Information and Electrionics, Beijing Institute of Technology, Beijing 100081, China. Li Shen is with the School of Cyber Science and Technology, Shen-zhen Campus of Sun Y at-sen University, Shenzhen 518107, China. Y ong Luo is with the School of Computer Science, National Engineering Research Center for Multimedia Software, Wuhan University, Wuhan 430072, China. Fuhui Sun and Xiaoyan Wang are with Information Technology Service Center of People's Court, Beijing, 100745, China. Y onggang Wen is with the College of Computing and Data Science, Nanyang Technological University, Singapore 639798. CoFormer significantly outperforms other methods. Specifically, CoFormer accelerates inference speed by 3.1 compared to Swin-L [4] with only 1.7% accuracy sacrifice.

Although traditional statistical techniques and machine learning methods have contributed significantly to genetics and, in particular, inherited disease diagnosis, they often struggle with complex, high-dimensional data, a challenge now addressed by state-of-the-art deep learning models. Large language models (LLMs), based on transformer architectures, have excelled in tasks requiring contextual comprehension of unstructured medical data. This systematic review examines the role of LLMs in the genetic research and diagnostics of both rare and common diseases. Automated keyword-based search in PubMed, bioRxiv, medRxiv, and arXiv was conducted, targeting studies on LLM applications in diagnostics and education within genetics and removing irrelevant or outdated models. A total of 172 studies were analyzed, highlighting applications in genomic variant identification, annotation, and interpretation, as well as medical imaging advancements through vision transformers. Key findings indicate that while transformer-based models significantly advance disease and risk stratification, variant interpretation, medical imaging analysis, and report generation, major challenges persist in integrating multimodal data (genomic sequences, imaging, and clinical records) into unified and clinically robust pipelines, facing limitations in generalizability and practical implementation in clinical settings. This review provides a comprehensive classification and assessment of the current capabilities and limitations of LLMs in transforming hereditary disease diagnostics and supporting genetic education, serving as a guide to navigate this rapidly evolving field.

Rapid and reliable qualification of advanced materials remains a bottleneck in industrial manufacturing, particularly for heterogeneous structures produced via non-conventional additive manufacturing processes. This study introduces a novel framework that links microstructure informatics with a range of expert characterization knowledge using customized and hybrid vision-language representations (VLRs). By integrating deep semantic segmentation with pre-trained multi-modal models (CLIP and FLAVA), we encode both visual microstructural data and textual expert assessments into shared representations. To overcome limitations in general-purpose embeddings, we develop a customized similarity-based representation that incorporates both positive and negative references from expert-annotated images and their associated textual descriptions. This allows zero-shot classification of previously unseen microstructures through a net similarity scoring approach. Validation on an additively manufactured metal matrix composite dataset demonstrates the framework's ability to distinguish between acceptable and defective samples across a range of characterization criteria. Comparative analysis reveals that FLAVA model offers higher visual sensitivity, while the CLIP model provides consistent alignment with the textual criteria. Z-score normalization adjusts raw unimodal and cross-modal similarity scores based on their local dataset-driven distributions, enabling more effective alignment and classification in the hybrid vision-language framework. The proposed method enhances traceability and interpretability in qualification pipelines by enabling human-in-the-loop decision-making without task-specific model retraining. By advancing semantic interoperability between raw data and expert knowledge, this work contributes toward scalable and domain-adaptable qualification strategies in engineering informatics.

The rapid development of artificial intelligence (AI), marked by breakthroughs like 'AlphaEvolve' and 'Gemini Deep Think', is beginning to offer powerful new tools that have the potential to significantly alter the research practice in many areas of mathematics. This paper explores the current landscape of publicly accessible large language models (LLMs) in a mathematical research context, based on developments up to August 2, 2025. Our analysis of recent benchmarks, such as MathArena and the Open Proof Corpus (Balunović et al., 2025; Dekoninck et al., 2025), reveals a complex duality: while state-of-the-art models demonstrate strong abilities in solving problems and evaluating proofs, they also exhibit systematic flaws, including a lack of self-critique and a model depending discrepancy between final-answer accuracy and full-proof validity. Based on these findings, we propose a durable framework for integrating AI into the research workflow, centered on the principle of the augmented mathematician. In this model, the AI functions as a copilot under the critical guidance of the human researcher, an approach distilled into five guiding principles for effective and responsible use. We then systematically explore seven fundamental ways AI can be applied across the research lifecycle, from creativity and ideation to the final writing process, demonstrating how these principles translate into concrete practice. We conclude that the primary role of AI is currently augmentation rather than automation. This requires a new skill set focused on strategic prompting, critical verification, and methodological rigor in order to effectively use these powerful tools.

AI-powered influence operations can now be executed end-to-end on commodity hardware. We show that small language models produce coherent, persona-driven political messaging and can be evaluated automatically without human raters. Two behavioural findings emerge. First, persona-over-model: persona design explains behaviour more than model identity. Second, engagement as a stressor: when replies must counter-arguments, ideological adherence strengthens and the prevalence of extreme content increases. We demonstrate that fully automated influence-content production is within reach of both large and small actors. Consequently, defence should shift from restricting model access towards conversation-centric detection and disruption of campaigns and coordination infrastructure. Paradoxically, the very consistency that enables these operations also provides a detection signature.

Retrieval-Augmented Generation (RAG) enhances large language models by incorporating external knowledge, yet suffers from critical limitations in high-stakes domains -- namely, sensitivity to noisy or contradictory evidence and opaque, stochastic decision-making. We propose ArgRAG, an explainable, and contestable alternative that replaces black-box reasoning with structured inference using a Quantitative Bipolar Argumentation Framework (QBAF). ArgRAG constructs a QBAF from retrieved documents and performs deterministic reasoning under gradual semantics. This allows faithfully explaining and contesting decisions. Evaluated on two fact verification benchmarks, PubHealth and RAGuard, ArgRAG achieves strong accuracy while significantly improving transparency.

Spiking Neural Networks (SNNs), inspired by the sparse event-driven neural processing and energy-efficient information transmission in the brain, offer a promising alternative for achieving low-power computing. Integrating the event-driven efficiency of spiking neurons with the advanced capabilities of LLMs represents a promising direction for power-efficient LLMs. This work specifically delves into the design of compressed spiking LLMs. Here, we revisit spatial and temporal pruning from the perspective of SNNs and propose a novel spatio-temporal pruning framework for Spiking LLMs to optimize computational efficiency while preserving high performance. Our spatial pruning technique reduces the number of active neurons and attention heads, effectively lowering the computational complexity of the model. Meanwhile, temporal pruning minimizes inference latency by dynamically adjusting the number of timesteps required for different layers. By combining these approaches with other compression techniques, we present the first work in the domain of Spiking LLMs to jointly explore spatial pruning, temporal pruning, extreme quantization and knowledge distillation strategies. Extensive experimental evaluation of our proposed framework for SpikingBERT on the large-scale GLUE benchmark demonstrates the efficacy of our approach in terms of computational operations and inference latency. Our approach offers a compelling solution for real-time, low-power natural language processing applications, making Spiking LLMs more practical for deployment on edge devices and in power-constrained settings.