Protein language models have emerged as powerful tools for sequence generation, offering substantial advantages in functional optimization and denovo design. However, these models also present significant risks of generating harmful protein sequences, such as those that enhance viral transmissibility or evade immune responses. These concerns underscore critical biosafety and ethical challenges. To address these issues, we propose a Knowledge-guided Preference Optimization (KPO) framework that integrates prior knowledge via a Protein Safety Knowledge Graph. This framework utilizes an efficient graph pruning strategy to identify preferred sequences and employs reinforcement learning to minimize the risk of generating harmful proteins. Experimental results demonstrate that KPO effectively reduces the likelihood of producing hazardous sequences while maintaining high functionality, offering a robust safety assurance framework for applying generative models in biotechnology.

Large language models (LLMs) often show poor performance in low-resource languages like Korean, partly due to unique linguistic challenges such as homophonous Sino-Korean words that are indistinguishable in Hangul script. To address this semantic ambiguity, we propose HanjaBridge, a novel meaning-injection technique integrated into a continual pre-training (CPT) framework. Instead of deterministically mapping a word to a single Hanja (Chinese character), HanjaBridge presents the model with all possible Hanja candidates for a given homograph, encouraging the model to learn contextual disambiguation. This process is paired with token-level knowledge distillation to prevent catastrophic forgetting. Experimental results show that HanjaBridge significantly improves Korean language understanding, achieving a 21\% relative improvement on the KoBALT benchmark. Notably, by reinforcing semantic alignment between Korean and Chinese through shared Hanja, we observe a strong positive cross-lingual transfer. Furthermore, these gains persist even when Hanja augmentation is omitted at inference time, ensuring practical efficiency with no additional run-time cost.

Recent advancements in dialogue generation have broadened the scope of human-bot interactions, enabling not only contextually appropriate responses but also the analysis of human affect and sensitivity. While prior work has suggested that stylistic similarity between user and system may enhance user impressions, the distinction between subjective and objective similarity is often overlooked. To investigate this issue, we introduce a novel dataset that includes users' preferences, subjective stylistic similarity based on users' own perceptions, and objective stylistic similarity annotated by third party evaluators in open-domain dialogue settings. Analysis using the constructed dataset reveals a strong positive correlation between subjective stylistic similarity and user preference. Furthermore, our analysis suggests an important finding: users' subjective stylistic similarity differs from third party objective similarity. This underscores the importance of distinguishing between subjective and objective evaluations and understanding the distinct aspects each captures when analyzing the relationship between stylistic similarity and user preferences. The dataset presented in this paper is available online.

Therapy recommendation for chronic patients with multimorbidity is challenging due to risks of treatment conflicts. Existing decision support systems face scalability limitations. Inspired by the way in which general practitioners (GP) manage multimorbidity patients, occasionally convening multidisciplinary team (MDT) collaboration, this study investigated the feasibility and value of using a Large Language Model (LLM)-based multi-agent system (MAS) for safer therapy recommendations. We designed a single agent and a MAS framework simulating MDT decision-making by enabling discussion among LLM agents to resolve medical conflicts. The systems were evaluated on therapy planning tasks for multimorbidity patients using benchmark cases. We compared MAS performance with single-agent approaches and real-world benchmarks. An important contribution of our study is the definition of evaluation metrics that go beyond the technical precision and recall and allow the inspection of clinical goals met and medication burden of the proposed advices to a gold standard benchmark. Our results show that with current LLMs, a single agent GP performs as well as MDTs. The best-scoring models provide correct recommendations that address all clinical goals, yet the advices are incomplete. Some models also present unnecessary medications, resulting in unnecessary conflicts between medication and conditions or drug-drug interactions.

Manipulation tasks are a key milestone toward integrating r obots into everyday life. They are particularly challenging because they require direct interac tion with objects, demanding higher levels of precision and robustness [ 1, 2 ]. Mobile manipulation - the integration of navigation and obj ect manipulation - is essential for a domestic service robot, as it enables a single platform to perf orm diverse tasks in unstructured home environments [ 3 ]. While fixed robots excel at repetitive factory operations, they lack the flexibility required to handle household chores. Consequently, the ability to generalize across tasks and environments becomes a critical capability for a truly vers atile domestic assistant. In recent years, learning-based control paradigms have gai ned significant traction for adapting to the unstructured environments encountered in manipulatio n tasks [ 4, 5 ]. Among these approaches, vision-based methods are especially appealing, as they lev erage the ease of obtaining visual data and enable end-to-end mapping from raw images to action outputs . These methods--directly generating robot joint motions from camera images--facilitate the rapi d deployment of domestic robots by eliminating the need for costly, large-scale sensor setups . However, end-to-end mapping has several drawbacks.

Real-time Automatic Speech Recognition (ASR) is a fundamental building block for many commercial applications of ML, including live captioning, dictation, meeting transcriptions, and medical scribes. Accuracy and latency are the most important factors when companies select a system to deploy. We present WhisperKit, an optimized on-device inference system for real-time ASR that significantly outperforms leading cloud-based systems. We benchmark against server-side systems that deploy a diverse set of models, including a frontier model (OpenAI gpt-4o-transcribe), a proprietary model (Deepgram nova-3), and an open-source model (Fireworks large-v3-turbo).Our results show that WhisperKit matches the lowest latency at 0.46s while achieving the highest accuracy 2.2% WER. The optimizations behind the WhisperKit system are described in detail in this paper.

Phishing remains a pervasive and growing threat, inflicting heavy economic and reputational damage. While machine learning has been effective in real-time detection of phishing attacks, progress is hindered by lack of large, high-quality datasets and benchmarks. In addition to poor-quality due to challenges in data collection, existing datasets suffer from leakage and unrealistic base rates, leading to overly optimistic performance results. In this paper, we introduce PhreshPhish, a large-scale, high-quality dataset of phishing websites that addresses these limitations. Compared to existing public datasets, PhreshPhish is substantially larger and provides significantly higher quality, as measured by the estimated rate of invalid or mislabeled data points. Additionally, we propose a comprehensive suite of benchmark datasets specifically designed for realistic model evaluation by minimizing leakage, increasing task difficulty, enhancing dataset diversity, and adjustment of base rates more likely to be seen in the real world. We train and evaluate multiple solution approaches to provide baseline performance on the benchmark sets. We believe the availability of this dataset and benchmarks will enable realistic, standardized model comparison and foster further advances in phishing detection. The datasets and benchmarks are available on Hugging Face (https://huggingface.co/datasets/phreshphish/phreshphish).

This paper demonstrates that Semantic Context (SC), leveraging descriptive tool information, is a foundational component for robust tool orchestration. Our contributions are threefold. First, we provide a theoretical foundation using contextual bandits, introducing SC-LinUCB and proving it achieves lower regret and adapts favourably in dynamic action spaces. Second, we provide parallel empirical validation with Large Language Models, showing that SC is critical for successful in-context learning in both static (efficient learning) and non-stationary (robust adaptation) settings. Third, we propose the FiReAct pipeline, and demonstrate on a benchmark with over 10,000 tools that SC-based retrieval enables an LLM to effectively orchestrate over a large action space. These findings provide a comprehensive guide to building more sample-efficient, adaptive, and scalable orchestration agents.

This paper introduces MISS-QA, the first benchmark specifically designed to evaluate the ability of models to interpret schematic diagrams within scientific literature. MISS-QA comprises 1,500 expert-annotated examples over 465 scientific papers. In this benchmark, models are tasked with interpreting schematic diagrams that illustrate research overviews and answering corresponding information-seeking questions based on the broader context of the paper. We assess the performance of 18 frontier multimodal foundation models, including o4-mini, Gemini-2.5-Flash, and Qwen2.5-VL. We reveal a significant performance gap between these models and human experts on MISS-QA. Our analysis of model performance on unanswerable questions and our detailed error analysis further highlight the strengths and limitations of current models, offering key insights to enhance models in comprehending multimodal scientific literature.

Individuals are increasingly relying on large language model (LLM)-enabled conversational agents for emotional support. While prior research has examined privacy and security issues in chatbots specifically designed for mental health purposes, these chatbots are overwhelmingly "rule-based" offerings that do not leverage generative AI. Little empirical research currently measures users' privacy and security concerns, attitudes, and expectations when using general-purpose LLM-enabled chatbots to manage and improve mental health. Through 21 semi-structured interviews with U.S. participants, we identified critical misconceptions and a general lack of risk awareness. Participants conflated the human-like empathy exhibited by LLMs with human-like accountability and mistakenly believed that their interactions with these chatbots were safeguarded by the same regulations (e.g., HIPAA) as disclosures with a licensed therapist. We introduce the concept of "intangible vulnerability," where emotional or psychological disclosures are undervalued compared to more tangible forms of information (e.g., financial or location-based data). To address this, we propose recommendations to safeguard user mental health disclosures with general-purpose LLM-enabled chatbots more effectively.