The following sources of funding for each author and for the preparation of this manuscript are gratefully acknowledged: J.Y.H.Y., J.K., P.Y. and Y.C. are supported by the AIR@innoHK programme of the Innovation and Technology Commission of Hong Kong. Additional support to J.Y.H.Y. was provided by the Judith and David Coffey Research Fund. Additional support to J.Y.H.Y. and Y.C was provided by the Chan Zuckerberg Initiative Single-Cell Biology Data Insights grant (DI2-0000000197). J.Y.H.Y., W.D., J.G. and X.L. are supported by the National Health and Medical Research Council (NHMRC) Investigator Grant (APP2017023). P.Y. is supported by an NHMRC Investigator Grant (1173469) and a Metcalf Prize from the National Stem Cell Foundation of Australia. J.K. is supported by the Australian Research Council Discovery Project (DP200103748). X.Z. is supported by the National Institutes of Health grant R01LM012976. D.K. is supported by an Australian Government Research Training Program Stipend Scholarship and the Children’s Medical Research Institute Top-Up Award. X.L. is supported by the University of Sydney Tuition Fee Scholarship. F.Y. is supported by the U.S. Department of Defense FY21 MRP Team Science Award (W81XWH-22-1-0731). R.J. is supported by the University of Sydney International Stipend Scholarship and Tuition Fee Scholarship. The funding sources had no role...

Artificial intelligence chatbots are rapidly gaining popularity for acquiring health information, but the quality of responses seems to differ between platforms. This study investigated the responses of ChatGPT, Gemini, Copilot, Claude, and DeepSeek to 15 frequently asked questions regarding femoral head avascular necrosis. Each platform answered each question once, for a total of 75 responses. Four orthopedic surgeons assessed the response using the JAMA benchmark criteria, DISCERN, an adapted five-point scale of clinical accuracy, and a binary safety score. Readability was measured by Flesch reading ease score (FRES) and Flesch-Kincaid grade level (FKGL). Models were compared by using Friedman tests and Bonferroni-corrected Wilcoxon signed-rank tests. Claude ranked the highest by JAMA and DISCERN scores and that DeepSeek outperformed ChatGPT, Gemini, and Copilot in terms of DISCERN. Clinical accuracy differed significantly among the models (p < 0.001). Claude and DeepSeek were rated higher than ChatGPT, Gemini, and Copilot, with no significant difference between Claude and DeepSeek. All responses were rated safe by all four raters. Readability performance scores of the models achieved significant differences on FRES (p = 0.043) and FKGL (p < 0.001) measures. The average FKGL score (10.85–13.93) was higher than the grade scores recommended for patient education information...

AbOptimizer is a policy gradient-based RL framework consisting of several components (Fig. 1b): agent, state, action, environment, and reward feedback. Given a starting sequence as input (initial state), the agent feeds it into a transformer-based policy network (referred to as the editor), which maps the state to an action space that marks each site for ‘reservation’ or ‘substitution’, while allowing ‘insertion’ and ‘deletion’ as optional sequence-editing choices. Then, the environment is constructed to interact with the agent and comprises an interpreter to modify the sequence/state based on the action taken, along with a Pareto-based reward function (referred to as evaluator) to score the quality of actions. The interpreter utilizes an antibody pre-trained LLM to predict plausible mutations for residues marked for ‘substitution’ and ‘insertion’. For ‘reservation’ sites, the corresponding residues are left unchanged, while for ‘deletion’ sites, they are pruned. The evaluator relies on well-validated predictors, called surrogates, which take arbitrary sequences as inputs and return their property scores. Through trial and error between the agent and the environment using a proximal policy optimization (PPO) algorithm, the policy network is ultimately trained to engineer variants with desired properties (Methods and Supplementary Section 1.1). We evaluated AbOptimizer under a tiered evaluation...

AbOptimizer is a policy gradient-based RL framework consisting of several components (Fig. 1b): agent, state, action, environment, and reward feedback. Given a starting sequence as input (initial state), the agent feeds it into a transformer-based policy network (referred to as the editor), which maps the state to an action space that marks each site for ‘reservation’ or ‘substitution’, while allowing ‘insertion’ and ‘deletion’ as optional sequence-editing choices. Then, the environment is constructed to interact with the agent and comprises an interpreter to modify the sequence/state based on the action taken, along with a Pareto-based reward function (referred to as evaluator) to score the quality of actions. The interpreter utilizes an antibody pre-trained LLM to predict plausible mutations for residues marked for ‘substitution’ and ‘insertion’. For ‘reservation’ sites, the corresponding residues are left unchanged, while for ‘deletion’ sites, they are pruned. The evaluator relies on well-validated predictors, called surrogates, which take arbitrary sequences as inputs and return their property scores. Through trial and error between the agent and the environment using a proximal policy optimization (PPO) algorithm, the policy network is ultimately trained to engineer variants with desired properties (Methods and Supplementary Section 1.1). We evaluated AbOptimizer under a tiered evaluation...

Lazer, D. et al. Computational social science. Science 323, 721–723 (2009). Lazer, D. M. J. et al. Computational social science: obstacles and opportunities. Science 369, 1060–1062 (2020). Vosoughi, S., Roy, D. & Aral, S. The spread of true and false news online. Science 359, 1146–1151 (2018). Bail, C. A. et al. Exposure to opposing views on social media can increase political polarization. Proc. Natl Acad. Sci. USA 115, 9216–9221 (2018). Kryvasheyeu, Y. et al. Rapid assessment of disaster damage using social media activity. Sci. Adv. 2, e1500779 (2016). Einav, L. & Levin, J. Economics in the age of big data. Science 346, 1243089 (2014). Ziems, C. et al. Can large language models transform computational social science? Comput. Linguist. 50, 237–291 (2024). Gabriel, I., Keeling, G., Manzini, A. & Evans, J. We need a new ethics for a world of AI agents. Nature 644, 38–40 (2025). Chen, L. et al. AI agent behavioral science. Humanit. Soc. Sci. Commun. 13, 1011 https://doi.org/10.1057/s41599-026-07316-7 (2026). Brinkmann, L. et al. Machine culture. Nat. Hum. Behav. 7, 1855–1868 (2023). Cheng, M., Piccardi, T. & Yang, D. CoMPosT: characterizing and evaluating caricature in LLM simulations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language...

We evaluated product name recommendations from five LLMs using 177,120 synthetic clinical vignettes generated by combining 123 curated disease–drug pairs with 1,440 prespecified sociodemographic profiles (Fig. 1 and Supplementary Table 1). Each pair linked a common disease to a prespecified active ingredient, and the profiles represented prespecified combinations of sociodemographic factors (see Methods for details). The five LLMs were GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Flash, Llama 4 and MedGemma. Using a standardized prompt, each LLM was asked to recommend up to three FDA-approved drugs by their product names. The five models generated a total of 883,403 responses with at least one drug recommendation (range per model: 175,197–177,119; Supplementary Fig. 1 and Supplementary Table 2). Among these recommendations, FDA-approved drugs corresponding to the specified active ingredient were identified and aggregated by disease–drug pair. The most frequently recommended products per LLM for 123 disease–drug pairs are shown in Supplementary Table 3. Product name convergence across vignettes and models For each of the 123 disease–drug pairs, we aggregated first-choice recommendations across all patient vignettes and calculated the share accounted for by the most frequently recommended product. Across pairs and models, this product name convergence was pronounced, with the top product capturing a...

Meta가 2026년 10월 15일 공개한 Muse Spark Open은 Mac에서 로컬 실행을 겨냥한 오픈웨이트 AI 모델입니다. 8160억 개 파라미터와 100만 토큰 문맥 창을 내세우지만, 실제 Mac에서 돌리려면 최소 1TB RAM이 필요합니다. 이 글은 발표 내용을 기준으로

Startup Reflection AI Inc. today made the weights of Beam, its first large language model, available on Hugging Face. The company also released technical resources that will enable developers to...

In daily pediatric practice, it has become increasingly common to meet families who have already explored the symptoms of their children using large language models (LLMs) before coming to the hospital.1,2,3 As this behavior becomes more widespread, pediatricians are often required to review these AI-generated statements, clarify inaccuracies, and support families to understand which parts of information are reliable, and which may be misleading. A similar shift is also evident in medical education. In our pediatric infectious disease clinics, we increasingly meet students, pediatrics residents, as well as pediatric infectious disease fellows, who consult LLMs while preparing for patient presentations or exams. They frequently use these tools to review symptoms, revisit disease summaries, or outline diagnostic steps before coming to the bedside.4,5 For many trainees, LLMs now function as a reliable source of clarification, offering explanations whenever they feel unsure.4,5 These shifts in both patient expectations and trainee learning habits naturally raised a question in our own clinical practice: how reliable and clinically appropriate these answers are provided by LLMs, particularly in pediatric infectious disease. This field requires age-specific interpretation and careful decision-making, and even small inaccuracies may influence the management.6 Recent comprehensive reviews have emphasized both the rapid integration...

The age-adjusted rate of drug-involved deaths in the US was ~250% higher than two decades ago1, despite a 24% reduction in overdose-related fatalities between 2023 and 20242. Adverse outcomes from overdoses extend beyond fatalities to other forms of measurable harm, including increased strain on the medical system. In 2024 alone the Drug Abuse Warning Network (DAWN) reported more than 8.4 million nonfatal drug-related emergency department (ED) visits in the US3. Meanwhile, the National Drug Early Warning System (NDEWS) identified parallel social media conversations around emerging drugs separately linked to increased ED visits in 2025, including synthetic opioids (e.g., nitazenes), opioid-like compounds (e.g., kratom), and polydrug-use combinations (e.g., benzodiazepines, kratom, or ketamine combined with fentanyl)4. Proactively addressing drug harm requires clinicians, health professionals, and policymakers to understand contemporary and evolving drug use trends as they emerge. Yet traditional epidemiologic surveillance systems (e.g., DAWN and others) contain built-in lags ranging from months to years5. Recommendations informed by epidemiologic patterns observed in data may occur even later. In addition, these systems, which typically measure prevalence and measurable impacts of drug use across communities and populations, similarly lack the capacity to capture new drug trends in places where they are likely to emerge, including...

To understand how LLMs perform in predicting data containing both coordinate and category information, we created synthetic datasets. Each data point consists of 3D coordinates (coordinate information, analogous to position of atoms) and an associated categorical label (analogous to atom types). These datasets are inspired by materials or molecular property-prediction tasks, where the objective is to predict the property from coordinates and atom types. We computed scalar target values for these coordinate–category data points using a physics-inspired hypothetical potential function (which can be thought of as an energy function). To that end, we construct the hypothetical potential such that we can continuously tune the impact of categories (that is, types) versus coordinate information (equation (1)). We give a more detailed explanation of equation (1) in ‘The physics-inspired hypothetical potential’ in the Methods. $$E=\alpha {E}_{\mathrm{category}}+(1-\alpha ){E}_{\mathrm{coordinate}}\,\alpha \in [0,1].$$ (1) In the extreme of α = 0, the target values depend solely on Ecoordinate (the energy contribution from the coordinate information, that is, the positions of the points in space), while for α = 1, the labels depend only on Ecategory (the energy contribution from the category information, that is, the types of points). On the basis of equation (1) and our...

Large-scale surveys are an important tool in economics. By directly eliciting latent preferences, they complement revealed-preference approaches and support both causal and descriptive inference at the population scale1,2. However, implementing such surveys is expensive and slow. Researchers need to construct sampling frames, train enumerators, provide respondent incentives and coordinate complex field operations. Recent advances in large language models (LLMs) have prompted an intriguing alternative: using LLMs to generate responses for surveys or experiments3,4,5,6,7,8,9,10. This approach can lower costs and accelerate iterations without human recruitment. Despite this promise, concerns remain about using LLMs’ simulated responses to replace human data, given inherent algorithm biases such as identity flattening11; systematic skew towards Western, educated, industrialized, rich and democratic populations12,13,14,15,16; overestimated statistical power7,17; and other potential risks18,19,20,21. Prior work has focused largely on political opinions6,17,22,23,24,25,26 and on social and moral attitudes12,13,14,27,28,29, while paying much less attention to economic preferences, even though these preferences play an important role in linking individual behaviour to market outcomes and social welfare, and constitute one of the most compelling use cases for LLMs in efforts to improve individual and societal well-being. Preference heterogeneity is a building block of modern economic theory: it helps explain individual decisions30,31,32, socio-economic status33,34 and...

Large language models (LLMs) have introduced a fundamental transformation in healthcare and medical education. This study aimed to evaluate the performance of LLMs in answering pharmacology questions. This quantitative cross-sectional study was conducted in Iran in 2026. A 50-item Persian multiple-choice questionnaire assessing pharmacology knowledge was administered to ChatGPT-5.5, Gemini 3.1, and Claude 4.6. Data were analyzed using confusion matrix methods, Cochran’s Q, McNemar, and point-biserial tests. Sensitivity values were 1.00 for ChatGPT-5.5, 0.94 for Gemini 3.1, and 0.90 for Claude 4.6. Specificity values were 0.43, 0.50, and 0.35, respectively. Accuracy was 0.82 for both ChatGPT-5.5 and Gemini 3.1, and 0.72 for Claude 4.6. Overall, the mean sensitivity, specificity, and accuracy across all LLMs were 0.95, 0.43, and 0.78, respectively. The analysis further revealed a statistically significant decline in performance of ChatGPT-5.5 and Gemini 3.1 when addressing quantitative questions (φ = −0.286, p < .05). LLMs have potential to support professionals in educational/knowledge contexts of pharmacology, provided ethical considerations are addressed, though further research is needed to better understand their performance and limitations. Rights and permissions Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in...

Combinatorial optimization (CO) underpins decision-making across science, engineering and industry, especially manufacturing, logistics, healthcare operations and energy management. However, designing effective heuristics for each new problem variant requires deep domain knowledge and months of expert iteration. Directly prompting large language models (LLMs) can generate executable code, but single-pass synthesis of complete heuristics routinely fails on new CO problems. Here we introduce LLM-driven Algorithm Construction via Complementary Evolution (LACE), a framework that decomposes algorithm design into an input schema, output schema, tool library and heuristic portfolio (the \({\mathcal{I}}\)–\({\mathcal{O}}\)\({\mathcal{T}}\)\({\mathcal{H}}\) interface) and refines the portfolio through time-constrained complementary evolution. The interface specifies a verified problem contract under which the LLM operates, directing model capacity to high-level algorithmic reasoning rather than problem-specific implementation. Complementary evolution then iteratively generates and selects specialist heuristics capable of solving heterogeneous instances within strict runtime budgets. Across 36 classical CO-Bench problems, LACE attains an average score of 0.945, against 0.870 for the strongest existing LLM-based method and 0.571 for direct LLM prompting without any framework, locating the gain in the framework rather than the model. On four structurally new problems, LACE reaches 0.97–0.99, while five existing LLM-based baselines fail to produce any feasible algorithm. These results demonstrate that effective...