Large Language Models (LLMs) are complicated artificial intelligence that are being used more in educational settings. Applying them into colleges and universities, particularly in mathematics-related fields, seems to be a good way to help students learn and get engaged. The objective of this study is to investigate undergraduate students' perceptions of the efficacy of LLMs in facilitating the understanding of mathematical concepts at selected universities in Malaysia. A quantitative survey design was used, involving undergraduate students enrolling in mathematics courses at various Malaysian higher education institutions. Data were gathered using structured questionnaires and processed employing descriptive and inferential statistical techniques. The results show that using LLMs has a big positive impact on how confident, motivated, and knowledgeable students are in math's. Respondents said that LLMs help them grasp abstract math issues better by giving them simple explanations, detailed problem-solving techniques, and specific feedback. Also, the fact that AI-generated help is there and easy to get to encourage students to study independently and makes them feel more comfortable in solving arithmetic problems. The study finds that LLMs are beneficial digital learning tools that help students learn more about math ideas at Malaysian institutions by centred on the students themselves.

With the increasing complexity and quantity of experimental and observational data, life scientists rely on programming to automate analyses, enhance reproducibility, and facilitate collaboration. Scripting languages like Python are often favored for their simplicity and flexibility, enabling researchers to focus primarily on high-level tasks. Compiled languages such as C++ and Rust offer greater efficiency, making them preferable for intensive or repeated computations. In educational settings, instructors may wish to teach both types of languages and thus may wish to translate content from one programming language to another. In research contexts, researchers may wish to implement their ideas in one language before translating the code to another. However, translating between programming languages requires significant effort, prompting our interest in using large language models (LLMs) for semi-automated code translation. This study explores the use of an LLM (GPT-4) to translate 559 short-form programming exercises from Python into C++, Rust, Julia, and JavaScript. We used three prompting strategies-instructions only, code only, or both combined-and compared the translated code's output against the Python code's output. Translation success differed considerably by prompting strategy, and at least one of the strategies tested was effective for nearly every exercise. The highest overall success rate occurred for Rust (99.5%), followed by JavaScript (98.9%), C++ (97.9%), and Julia (95.0%). Our findings demonstrate that LLMs can effectively translate small-scale programming exercises between languages, reducing the need for manual rewriting. To support education and research, we have manually translated all exercises that were not translated successfully through automation, and we have made our translations freely available.

From Corpus to Innovation: Advancing Organic Solar Cell Design with Large Language Models

The Congress of Neurological Surgeons Self-Assessment for Neurological Surgeons questions are widely used by neurosurgical residents to prepare for written board examinations. Recently, these questions have also served as benchmarks for evaluating large language models' (LLMs) neurosurgical knowledge. LLMs show significant promise for transforming neurosurgical practice; however, they are susceptible to in-text distractions and confounding factors. Given the increasing use of generative artificial intelligence and ambient dictation technologies, clinical text is at a larger risk for the inclusion of extraneous details. The aim of this study was to assess the performance of state-of-the-art LLMs on neurosurgery board-like questions and to evaluate their robustness to the inclusion of distractor statements. A comprehensive evaluation was conducted using 28 state-of-the-art LLMs. These models were tested on 2904 neurosurgery board examination questions derived from the Congress of Neurological Surgeons Self-Assessment for Neurological Surgeons. In addition, the study introduced a distraction framework to assess the fragility of these models. The framework incorporated simple, irrelevant distractor statements containing polysemous words with clinical meanings used in nonclinical contexts to determine the extent to which such distractions degrade model performance on standard medical benchmarks. Six of the 28 tested LLMs achieved board-passing outcomes, with the top-performing models scoring over 15.7% above the passing threshold. When exposed to distractions, accuracy across various model architectures was significantly reduced-by as much as 20.4%-with 1 model failing that had previously passed. Both general-purpose and medical open-source models experienced greater performance declines compared with proprietary variants when subjected to the added distractors. While current LLMs demonstrate an impressive ability to answer neurosurgery board-like examination questions, their performance is markedly vulnerable to extraneous, distracting information. These findings underscore the critical need for developing novel mitigation strategies aimed at bolstering LLM resilience against in-text distractions, particularly for safe and effective clinical deployment.

PurposeTo evaluate the utility of CataractBot, an LLM (Large Language Model)-powered chatbot that provides doctor-verified answers to patient questions about cataract surgery. We examine its use by both end-users (patients and attendants) and medical experts.MethodsA 24-week study was conducted to evaluate CataractBot among patients, their attendants, doctors, and patient coordinators. The bot responded instantly to questions by querying a knowledge base curated by medical professionals. Each response was asynchronously verified by an ophthalmologist (for medical questions) or a patient coordinator (for logistical questions), and their edits contributed to updating the knowledge base, thereby minimizing future expert intervention. A mixed-methods analysis was conducted on interaction logs, including patient and attendant questions, chatbot answers, and expert verifications.ResultsA total of 318 patients and attendants sent 1,992 messages, and LLM-generated answers were verified by five doctors and two coordinators. Questions asked pre-surgery were significantly more than post-surgery . Participants asked significantly more medical than logistical questions . Doctors rated 84.5% of CataractBot's answers to medical questions as accurate and complete. Their edits, which mainly involved adding information, increased the acceptance of the bot's answers by 19.0% over time.ConclusionCataractBot was predominantly used to address medical questions. It incorporated expert corrections to improve its answers and reduce the experts' bot-related workload over time. This study highlights the potential of LLM-powered chatbots to support patient-provider communication in ophthalmology.

Generative artificial intelligence, particularly large language models (LLMs), is increasingly used to navigate information, potentially shaping users' perceptions of different social groups. This study examines age-related stereotypes in LLM-generated text using natural language processing (NLP) techniques. To ensure neutrality, extensive pilot testing was conducted to craft a prompt that did not elicit bias yet generated coherent responses. The final prompt, "Describe the personality of a [AGE]-year-old person," was used with OpenAI's GPT-4o API in February 2025, varying AGE from 10 to 90 in 10-year increments. The analysis was guided by the Stereotype Content Model, which assesses social cognition along two key dimensions: warmth (sociability, morality) and competence (ability, assertiveness). Scores were quantified using sentence embeddings. Text similarity and stereotype content analyses revealed three age clusters, with older adults showing the most internal consistency. Descriptions of individuals aged 60 and above featured relatively higher warmth but lower competence compared to younger groups. Notably, positive assertiveness terms were rarely used to describe older adults. Findings suggest that GPT-4o may embed subtle age-related stereotypes, even when using largely positive language. These patterns potentially influence user perceptions through repeated exposure. Future research should investigate the mechanisms behind these biases and explore mitigation strategies to promote more age-inclusive AI-generated content.

Large language models (LLMs) such as ChatGPT have gained attention for their potential to assist clinical decision-making in oncology. However, real-world validation of these models against multidisciplinary tumor board (MTB) recommendations-particularly in breast cancer treatment-remains limited.This retrospective study assessed the concordance between GPT-4o and the decisions of a breast cancer MTB over a six-month period. Thirty-three patients were included. Structured clinical data were entered into GPT-4o using standardized prompts, and treatment plans were generated in two independent sessions per case. Seven therapeutic domains were evaluated: surgery, radiotherapy, hormonal therapy, neoadjuvant therapy, adjuvant therapy, genetic counseling/testing, and dual HER2-targeted therapy. Two blinded reviewers scored concordance using a 5-point Likert scale. Inter-rater reliability and classification metrics were calculated.GPT-4o generated consistent recommendations across both sessions for all patients. Full concordance (5/5) with MTB decisions was observed in 31 of 33 cases (93.9%), while partial concordance (4/5) occurred in 2 cases (6.1%) due to differences regarding genetic counseling. Inter-rater agreement was perfect (Cohen's kappa = 1.00), and the mean concordance score was 4.94 out of 5. The model achieved an overall accuracy of 93.9%, precision of 93.9%, recall of 100%, and F1 score of 96.8%.GPT-4o demonstrated a high level of agreement with expert multidisciplinary decisions in breast cancer care when provided with structured clinical input. These findings support its potential as a reproducible, guideline-consistent decision-support tool in oncology workflows.

Purpose This paper introduces MedMaintBot, an AI chatbot designed to support biomedical technicians and non-expert users like nurses. The study explores the impact of integrating such an AI chatbot into Total Productive Maintenance (TPM) practices in healthcare, aligned with Industry 5.0 (I5.0) principles. Design/methodology/approach This study adopts a multi-phase methodology, starting with a literature review on technology integration in TPM within healthcare settings. It presents the chatbot development pipeline and conducts a large-scale validation study across 250 queries covering five medical devices (MDs) to demonstrate the chatbot's real-time, context-aware guidance capabilities. Performance analysis further evaluates MedMaintBot's potential to optimize TPM practices and support sustainability goals in healthcare maintenance. Findings The study reveals that MedMaintBot enhances TPM within healthcare by delivering accurate, context-aware guidance (Accuracy = 0.713, Relevance = 0.810), supporting nurse autonomy in routine maintenance and reducing technician dependency. While clarity and completeness were slightly below ideal for complex tasks, over 80% of autonomy-related queries were validated, showing strong support for first-level interventions. Combined with dynamic Large Language Model (LLM) switching between GPT-4 and MedLLaMA2, MedMaintBot strikes a balance between performance, cost and privacy, positioning it as a scalable and sustainable tool for healthcare maintenance. Research limitations/implications This research provides valuable insights for practitioners and researchers on enhancing autonomous maintenance (AM) through AI–chatbot integration, offering a scalable framework for integrating AI into TPM practices. It also encourages further studies to address gaps in procedural completeness and contextual continuity and assess scalability across diverse maintenance environments. Practical implications By providing real-time, context-aware guidance, the chatbot helps reduce user-induced errors, allowing non-expert users, such as nurses, to perform maintenance tasks. This not only reduces the burden on specialized technicians but also ensures better equipment availability, contributing to more streamlined healthcare operations and improved patient care. Social implications MedMaintBot promotes a more inclusive and resilient healthcare environment by empowering non-expert users with AI-driven support. Its adaptability aligns with the human-centric principles of Industry 5.0, fostering collaboration between technology and healthcare personnel. Originality/value This research is among the first to examine the integration of innovative AI chatbot with TPM practices within the healthcare sector, particularly in the context of I5.0. It demonstrates how such a system can significantly enhance operational efficiency, empower non-expert users and support sustainability in healthcare, offering a roadmap extending AI-assisted maintenance to broader industrial and resource-constrained environments.

Response to: "Several Key Methodological Issues Concerning the Application of Large Language Models in DISE Interpretation".

The discovery and engineering of new enzymes is important across the bioeconomy, with diverse applications from foods to pharmaceuticals, sensors to agriculture. However, enzyme engineering, in particular machine learning-guided engineering, is hampered by a lack of data. Currently there exists no database designed to capture and interpret datasets created in this domain, nor are there easy analysis and visualisation tools. We developed the Enzyme Engineering Database to provide a centralized resource and an online analysis tool to consolidate sequence-function data from enzyme engineering campaigns, thereby making three contributions: (i) a database into which researchers can deposit public data, (ii) visualisation and analysis tools for protein engineers to analyse their own data or compare enzyme variants to other engineering campaigns, and (iii) a gold-standard dataset for benchmarking automated extraction along with the first large language model extraction pipeline specific for enzyme engineering campaigns. The Enzyme Engineering Database is accessible at http://enzengdb.org/.

Comparative evaluation of large language models on multiple-choice and image-based rheumatology questions.

Background The European Union’s Artificial Intelligence Act (AI Act) establishes a pioneering, sector-agnostic, risk-based taxonomy for artificial intelligence systems. However, its concrete operational implications for academic libraries, which increasingly rely on AI for discovery services, metadata generation, user support, and analytics, remain insufficiently explored. Methods This study conducts a systematic review of recent scientific and professional literature and maps common academic library AI use cases onto the proportional risk and compliance obligations defined by the AI Act. Based on this analysis, a sector-specific risk-classification matrix is developed to support regulatory interpretation in the library context. Results The findings indicate that ethical principles frequently prevail over enforceable compliance mechanisms, that library AI applications align conceptually with the AI Act’s taxonomy but lack practical operationalization, that generative AI and large language models intensify regulatory ambiguity, that AI procurement practices in libraries rarely incorporate AI Act safeguards, and that the protection of fundamental rights, including equity, privacy, and intellectual freedom, requires measurable controls beyond transparency notices. Discussion The results reveal a substantial gap between the regulatory framework and its practical implementation in academic libraries. Governance approaches remain largely normative, while auditable and operational compliance mechanisms are still underdeveloped. The rapid diffusion of generative AI further complicates accountability, risk classification, and institutional responsibility. Conclusion This study contributes a sector-specific AI risk-classification matrix, identifies policy needs for tailored audit and procurement models, and highlights key research gaps in empirical validation, bias detection, and trust frameworks. By bridging regulation and practice, it positions academic libraries as potential norm-setters within the European AI governance ecosystem, exemplifying rights-preserving and compliance-ready institutions where regulation acts both as a safeguard and a catalyst for responsible innovation.

CASSIA: a multi-agent large language model for automated and interpretable cell annotation

Introduction: Large language models (LLMs) like ChatGPT are used by medical trainees and professionals for learning and clinical support. This study determined how Canadian plastic surgery residents utilize and perceive LLMs for their training. Methods: A cross-sectional survey was distributed to all Canadian, English-speaking plastic surgery trainees (N = 100). Descriptive statistics and conventional content analysis were used to describe quantitative and free-text responses, respectively. Results: A total of n = 36 responses were collected (36% response rate) from Canadian plastic surgery residents. Among residents, 83.3% reported using LLMs for any purpose, and 63.8% reported using the technology for plastic surgery education. The most frequently utilized LLMs include ChatGPT (83.3%), BingAI (11.1%), and Gemini (8.3%). More than half of residents reported using LLMs a minimum of once per week (50.1%). The most common applications included explaining concepts (58.3%), explaining procedures (33.3%), answering lecture questions (27.8%), and creating presentations (27.8%). Of respondents, 94.4% reported not having received education or training on the use of LLMs, and 37.1% reported concerns with the use of the technology for plastic surgery learning. The themes that emerged from the free-text responses were categorized into 3 groups: (1) advantages, including time-efficiency and summarization, (2) disadvantages, including concerns of inaccuracies, confidentiality, and over-reliance, and (3) recommendations, such as didactic teaching sessions and workshops. Conclusions: LLMs are commonly used by Canadian plastic surgery residents for a variety of purposes. Most residents have not been trained on the optimal use of the technology, and surgical residency programs should consider formal LLM instruction to leverage the capabilities of this tool and mitigate potential harms.

Staging Prostate Cancer with AI: A Comparative Study of Large Language Models and Expert Interpretation on PSMA PET-CT Reports.

Large language models (LLMs) are increasingly applied in health care. However, concerns remain that their nursing care recommendations may reflect patients' sociodemographic attributes rather than clinical needs. While this risk is acknowledged, there is a lack of empirical evidence evaluating sociodemographic bias in LLM-generated nursing care plans. To investigate potential biases in nursing care plans generated by LLMs, we focused on whether outputs differ systematically based on patients' sociodemographic characteristics and assessed the implications for equitable nursing care. We used a mixed methods simulation study. A standardized clinical vignette experiment was used to prompt GPT-4 to generate 9600 nursing care plans for 96 patient profiles with varying sociodemographic characteristics (eg, sex, age, income, education, and residence). We first conducted a quantitative analysis of all plans, assessing variations in thematic content. Subsequently, a panel of senior nursing experts evaluated the clinical quality (eg, safety, applicability, and completeness) of a stratified subsample of 500 plans. We analyzed 9600 LLM-generated nursing care plans and identified 8 consistent themes. Communication and Education (99.98%) and Emotional Support (99.97%) were nearly universal, while Nurse Training and Event Analysis were least frequent (39.3%). Multivariable analyses revealed systematic sociodemographic disparities. Care plans generated for low-income patient profiles were less likely to include the theme Environmental Adjustment (adjusted relative risk [aRR] 0.90). Profiles with lower education were associated with an increased likelihood of including Family Support (aRR 1.10). Similarly, plans generated for older patient profiles were more likely to contain recommendations for Pain Management (aRR 1.33) and Family Support (aRR 1.62) but were less likely to mention Nurse Training (aRR 0.78). Sex and regional differences were also significant. Expert review of 500 plans showed high overall quality (mean 4.47), with strong interrater reliability (κ=0.76-0.81). However, urban profiles had higher completeness (β=.22) and applicability (β=.14) but lower safety scores (β=-0.09). These findings demonstrate that LLM-generated care plans exhibit systematic sociodemographic bias, raising important implications for fairness and safe deployment in nursing practice. This study identified that LLMs systematically reproduce sociodemographic biases in the generation of nursing care plans. These biases appear in two forms: they shape the thematic content and influence expert-rated clinical quality. These findings reveal a substantial risk that such models may reinforce existing health inequities. To our knowledge, this is the first empirical evidence documenting these nuanced biases in nursing. The study also contributes a replicable framework for evaluating LLM-generated care plans. Finally, it underscores the critical need for robust human oversight to ensure that artificial intelligence serves as a tool for advancing equity rather than perpetuating disparities.

Introduction: Mass casualty incident (MCI) simulation scenarios are developed based on detailed review and planning by multidisciplinary trauma teams. This study aimed to assess the feasibility of using generative artificial intelligence (AI) in developing mass casualty trauma simulation scenarios. The study evaluated a range of mass casualty trauma simulation scenarios generated from a public generative artificial intelligence platform based on publicly available data with a validated objective simulation scoring tool. Methods: Using a large language model (LLM) platform (ChatGPT4, OpenAI, San Francisco, CA, USA), 10 complex MCI trauma simulation scenarios were generated based on publicly available US reported trauma data. Each scenario was evaluated by two Advanced Trauma Life Support (ATLS) certified raters based on the Simulation Scenario Evaluation Tool (SSET), a validated scoring tool out of 100 points. The tool scoring is based on learning objectives, tasks for performance, clinical progression, debriefing criteria, and resources. Two publicly available mass casualty trauma scenarios were similarly evaluated as controls. Revision and recommended feedback was provided for the scenarios, with review time recorded. Post-revision scenarios were evaluated. Interrater reliability was calculated based on Intraclass Correlation Coefficients (2, k) (ICCs). For the scenarios, scores and review times were reported as medians with interquartile range (IQR) as 25th and 75th percentiles. Results: Ten mass casualty trauma simulation scenarios were generated by an LLM, producing a total of 62 simulated patients. The initial LLM-generated scenarios demonstrated a median SSET score of 78.5 (IQR 74–82), substantially lower than the median score of 94 (IQR 93–95) observed in publicly available scenarios. The interrater reliability ICC for the LLM-generated scenarios was 0.965 and 1.00 for publicly available scenarios. Following secondary human revision and iterative refinement, the LLM-generated scenarios improved, achieving a median SSET score of 94 (IQR 93–96) with an interrater reliability ICC of 0.7425. Conclusions: The feasibility study suggests that a structured, collaborative workflow combining LLM-based generation with expert human review may enable a new approach to mass casualty trauma simulation scenario creation. LLMs hold promise as a scalable tool for the development of MCI training materials. However, consistent human oversight, quality assurance processes, and governance frameworks remain essential to ensure clinical accuracy, safety, and educational value.

BackgroundArtificial intelligence (AI) is reshaping the landscape of medical education with unprecedented depth and breadth. As technologies like large language models and natural language processing advance, AI agents with multimodal interaction capabilities—such as intelligent teaching assistants and virtual simulation labs—are demonstrating immense potential. Concurrently, medical students face challenges including a disconnect between theoretical knowledge and clinical practice, excessive cognitive load, and a lack of personalized practical opportunities. Medical education AI agents are poised to address these issues, but their successful integration hinges on student acceptance and adoption. This study aims to fill a gap in the current empirical research by investigating the key psychological mechanisms and behavioral factors that influence medical students’ adoption of AI educational agents.MethodsThis study constructed an extended Unified Theory of Acceptance and Use of Technology (UTAUT) model by integrating four key variables tailored to the medical education context: AI Trust, Perceived Risk, Hedonic Motivation, and Trialability. A cross-sectional survey was conducted with an initial sample of 200 clinical medicine students following their interaction with a custom-developed interactive medical education AI agent. After excluding invalid responses, a final valid sample of 155 participants was retained. Partial Least Squares Structural Equation Modeling (PLS-SEM) was employed to validate the theoretical model and test the research hypotheses.ResultsThe constructed model demonstrated strong explanatory power, successfully accounting for 85.3% of the variance in students’ behavioral intention (R² = 0.853). Effort Expectancy (β = 0.362, p < 0.001) and Performance Expectancy (β = 0.297, p < 0.001) were the strongest direct positive predictors of behavioral intention, with Facilitating Conditions (β = 0.258, p = 0.002) also showing a significant impact. A noteworthy finding was that Social Influence had no significant effect on behavioral intention (β = 0.038, p = 0.633). Furthermore, Hedonic Motivation had a significant positive influence on both Effort Expectancy (β = 0.818, p < 0.001) and Performance Expectancy (β = 0.237, p < 0.001). AI Trust, Trialability, and lower Perceived Risk also significantly enhanced students’ Performance Expectancy.ConclusionsThe findings indicate that for medical students, who are highly autonomous professional learners, the intrinsic value of an AI educational tool (i.e., its utility and ease of use) is the dominant factor in their adoption decisions, far outweighing the social influence of peers or authorities. Therefore, the key to successfully promoting such technologies lies in building users’ intrinsic trust, reducing their perceived risks, and providing an engaging, immersive learning experience. These findings provide a solid empirical basis for the optimal design of medical AI educational agents and for strategies to effectively integrate them into the curriculum.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12909-025-08234-z.

Large language model (LLM)-based artificial intelligence (AI) agents are powerful tools that can help researchers automate complex tasks such as literature review, data mining, computational code generation, and summarization of existing knowledge, but they can still fall short in developing original biological hypotheses and insights (see related Research Article by Rodriguez-Coffinet etal. in this issue). Emerging advances in multiagent systems and human-agent collaborative frameworks offer promising steps forward.

MediAstra is an integrated AI-powered healthcare ecosystem designed to unify medical assistance, personalized fitness management, and NEET exam preparation within a single intelligent application. Today, users rely on multiple fragmented platforms to access medical guidance, fitness planning, and educational support, leading to inefficiency and lack of comprehensive personalization. MediAstra overcomes this challenge by integrating machine learning, natural language processing, and rule-based decision models to deliver unified support across healthcare, wellness, and academic domains.The platform comprises four core modules: SymptoScan AI for symptom-based disease prediction, Fit360 AI for personalized fitness and diet planning, CareSync for healthcare navigation and consultation, and NeetMind AI for adaptive learning and NEET exam preparation. The system processes user inputs through secure authentication, validates data at the server level, interprets queries using the Groq LLM (Large Language Model), and stores interactions in a robust relational database for personalized insights. Keywords: