Tokenization inefficiency imposes structural disadvantages on morphologically complex, low-resource languages, inflating compute resources and depressing accuracy. We evaluate 10 large language models (LLMs) on AfriMMLU (9,000 MCQA items; 5 subjects; 16 African languages) and show that fertility (tokens/word) reliably predicts accuracy. Higher fertility consistently predicts lower accuracy across all models and subjects. We further find that reasoning models (DeepSeek, o1) consistently outperform non-reasoning peers across high and low resource languages in the AfriMMLU dataset, narrowing accuracy gaps observed in prior generations. Finally, translating token inflation to economics, a doubling in tokens results in quadrupled training cost and time, underscoring the token tax faced by many languages. These results motivate morphologically aware tokenization, fair pricing, and multilingual benchmarks for equitable natural language processing (NLP).

Hospital call centers serve as the primary contact point for patients within a hospital system. They also generate substantial volumes of staff messages as navigators process patient requests and communicate with the hospital offices following the established protocol restrictions and guidelines. This continuously accumulated large amount of text data can be mined and processed to retrieve insights; however, traditional supervised learning approaches require annotated data, extensive training, and model tuning. Large Language Models (LLMs) offer a paradigm shift toward more computationally efficient methodologies for healthcare analytics. This paper presents a multi-stage LLM-based framework that identifies staff message topics and classifies messages by their reasons in a multi-class fashion. In the process, multiple LLM types, including reasoning, general-purpose, and lightweight models, were evaluated. The best-performing model was o3, achieving 78.4% weighted F1-score and 79.2% accuracy, followed closely by gpt-5 (75.3% Weighted F1-score and 76.2% accuracy). The proposed methodology incorporates data security measures and HIPAA compliance requirements essential for healthcare environments. The processed LLM outputs are integrated into a visualization decision support tool that transforms the staff messages into actionable insights accessible to healthcare professionals. This approach enables more efficient utilization of the collected staff messaging data, identifies navigator training opportunities, and supports improved patient experience and care quality.

Large Language Models (LLMs) are increasingly vulnerable to a sophisticated form of adversarial prompting known as camouflaged jailbreaking. This method embeds malicious intent within seemingly benign language to evade existing safety mechanisms. Unlike overt attacks, these subtle prompts exploit contextual ambiguity and the flexible nature of language, posing significant challenges to current defense systems. This paper investigates the construction and impact of camouflaged jailbreak prompts, emphasizing their deceptive characteristics and the limitations of traditional keyword-based detection methods. We introduce a novel benchmark dataset, Camouflaged Jailbreak Prompts, containing 500 curated examples (400 harmful and 100 benign prompts) designed to rigorously stress-test LLM safety protocols. In addition, we propose a multi-faceted evaluation framework that measures harmfulness across seven dimensions: Safety Awareness, Technical Feasibility, Implementation Safeguards, Harmful Potential, Educational Value, Content Quality, and Compliance Score. Our findings reveal a stark contrast in LLM behavior: while models demonstrate high safety and content quality with benign inputs, they exhibit a significant decline in performance and safety when confronted with camouflaged jailbreak attempts. This disparity underscores a pervasive vulnerability, highlighting the urgent need for more nuanced and adaptive security strategies to ensure the responsible and robust deployment of LLMs in real-world applications.

Membership inference attacks (MIAs) reveal whether specific data was used to train machine learning models, serving as important tools for privacy auditing and compliance assessment. Recent studies have reported that MIAs perform only marginally better than random guessing against large language models, suggesting that modern pre-training approaches with massive datasets may be free from privacy leakage risks. Our work offers a complementary perspective to these findings by exploring how examining LLMs' internal representations, rather than just their outputs, may provide additional insights into potential membership inference signals. Our framework, \emph{memTrace}, follows what we call \enquote{neural breadcrumbs} extracting informative signals from transformer hidden states and attention patterns as they process candidate sequences. By analyzing layer-wise representation dynamics, attention distribution characteristics, and cross-layer transition patterns, we detect potential memorization fingerprints that traditional loss-based approaches may not capture. This approach yields strong membership detection across several model families achieving average AUC scores of 0.85 on popular MIA benchmarks. Our findings suggest that internal model behaviors can reveal aspects of training data exposure even when output-based signals appear protected, highlighting the need for further research into membership privacy and the development of more robust privacy-preserving training techniques for large language models.

As large-language models have been increasingly used as automatic raters for evaluating free-form content, including document summarization, dialog, and story generation, work has been dedicated to evaluating such models by measuring their correlations with human judgment. For \textit{sample-level} performance, methods which operate by using pairwise comparisons between machine-generated text perform well but often lack the ability to assign absolute scores to individual summaries, an ability crucial for use cases that require thresholding. In this work, we propose a direct-scoring method which uses synthetic summaries to act as pairwise machine rankings at test time. We show that our method performs comparably to state-of-the-art pairwise evaluators in terms of axis-averaged sample-level correlations on the SummEval (\textbf{+0.03}), TopicalChat (\textbf{-0.03}), and HANNA (\textbf{+0.05}) meta-evaluation benchmarks, and release the synthetic in-context summaries as data to facilitate future work.

Automating the conversion of user interface design into code (image-to-code or image-to-UI) is an active area of software engineering research. However, the state-of-the-art solutions do not achieve high fidelity to the original design, as evidenced by benchmarks. In this work, I approach the problem differently: I use vector images instead of bitmaps as model input. I create several large datasets for training machine learning models. I evaluate the available array of Image Quality Assessment (IQA) algorithms and introduce a new, multi-scale metric. I then train a large open-weights model and discuss its limitations.

Abstract: The use of large language models (LLMs) in bioethical, scientific, and medical writing remains controversial. While there is broad agreement in some circles that LLMs cannot count as authors, there is no consensus about whether and how humans using LLMs can count as authors. In many fields, authorship is distributed among large teams of researchers, some of whom -- including paradigmatic "senior authors" who guide and determine the scope of a project and ultimately vouch for its integrity -- may not write a singl e word. In this paper, we argue that LLM use (under specific conditions) is analogous to a form of senior authorship. On this view, the use of LLMs, even to generate complete drafts of research papers, can be considered a legitimate form of authorship according to the accepted criteria in many fields. We conclude that either such use should be recognized as legitimate, or current criteria for authorship require fundamental revision. AI use declaration: Chat GPT version 5 was used to help format Box 1. AI wa s not used for any other part of the preparation or writing of this manuscript. This is a pre print of a paper that has been submitted to a journal. It has not yet gone through peer review. Authorship Without Writing: Large Language Models and the "Senior Author" Analogy Clint Hurshman, Sebastian Porsdam Mann, Julian Savulescu, Brian D. Earp I. Introduction The use of large language models (LLMs) in bioethics as well as scientific and medical writing continues to be controversial. Thus far, there has been broad agreement -- for example, among medical publishers -- that LLMs cannot count as authors, but there is still no consensus about the status of LLM - assisted text production as a form of writing, and by extension, the status of LLM users as authors. Here, we contribute to this debate by exploring -- and drawing analogies to -- the collaborative nature of writing, and t he distributed character of authorship, in many domains of research.

Large language models (LLMs) are unable to continuously adapt and learn from new data during reasoning at inference time. To address this limitation, we propose that complex reasoning tasks be decomposed into atomic subtasks and introduce SAGE, a trigger-guided dynamic fine-tuning framework that enables adaptive updates during reasoning at inference time. SAGE consists of three key components: (1) a Trigger module that detects reasoning failures through multiple evaluation metrics in real time; (2) a Trigger Buffer module that clusters anomaly samples using a streaming clustering process with HDBSCAN, followed by stability checks and similarity-based merging; and (3) a Lora Store module that dynamically optimizes parameter updates with an adapter pool for knowledge retention. Evaluation results show that SAGE demonstrates excellent accuracy, robustness, and stability on the atomic reasoning subtask through dynamic knowledge updating during test time.

Model developers are eager to acquire new sources of high-quality training data as they race to improve model capabilities and win market share. This paper analyzes the privacy policies of six U.S. frontier AI developers to understand how they use their users' chats to train models. Drawing primarily on the California Consumer Privacy Act, we develop a novel qualitative coding schema that we apply to each developer's relevant privacy policies to compare data collection and use practices across the six companies. We find that all six developers appear to employ their users' chat data to train and improve their models by default, and that some retain this data indefinitely. Developers may collect and train on personal information disclosed in chats, including sensitive information such as biometric and health data, as well as files uploaded by users. Four of the six companies we examined appear to include children's chat data for model training, as well as customer data from other products. On the whole, developers' privacy policies often lack essential information about their practices, highlighting the need for greater transparency and accountability. We address the implications of users' lack of consent for the use of their chat data for model training, data security risks arising from indefinite chat data retention, and issues with training on children's chat data. We conclude by providing recommendations to policymakers and developers to address the data privacy challenges posed by LLM-powered chatbots.

As our cities and communities become smarter, the systems that keep us safe, such as traffic control centers, emergency response networks, and public transportation, also become more complex. With this complexity comes a greater risk of security threats that can affect not just machines but real people's lives. To address this challenge, we present ThreatGPT, an agentic Artificial Intelligence (AI) assistant built to help people whether they are engineers, safety officers, or policy makers to understand and analyze threats in public safety systems. Instead of requiring deep cybersecurity expertise, it allows users to simply describe the components of a system they are concerned about, such as login systems, data storage, or communication networks. Then, with the click of a button, users can choose how they want the system to be analyzed by using popular frameworks such as STRIDE, MITRE ATT&CK, CVE reports, NIST, or CISA. ThreatGPT is unique because it does not just provide threat information, but rather it acts like a knowledgeable partner. Using few-shot learning, the AI learns from examples and generates relevant smart threat models. It can highlight what might go wrong, how attackers could take advantage, and what can be done to prevent harm. Whether securing a city's infrastructure or a local health service, this tool adapts to users' needs. In simple terms, ThreatGPT brings together AI and human judgment to make our public systems safer. It is designed not just to analyze threats, but to empower people to understand and act on them, faster, smarter, and with more confidence.