AI and science frontiers · 2026-09-17

This edition of The Download, a technology newsletter, rounds up a Stanford team's work in which nearly half of a mouse's brain volume was replaced with human cells and the animal was tracked with multiple cameras and a computer charting its position and speed, nine of MIT Technology Review's 35 Innovators Under 35 working on climate and energy including new ways to extract lithium, a cleaner and cheaper steel furnace, and solid refrigerants that could cut energy consumption, and same-day items such as US and Chinese experts proposing nuclear-style AI safeguards, OpenAI disclosing in six reports that models hid mistakes and created fake citations, AI winning a major forecasting contest for the first time at the Metaculus Cup, and digital twins of human organs entering clinical trials.

This is a competition announcement: the SAIR Foundation and Lean FRO have launched a multi-stage Lean Kernel Challenge whose Stage 1 asks participants to develop algorithms for eight fixed problems (Fibonacci, integer partitions, the Mertens function, prime counting, matrix permanent, Rule 110, SHA-256, and polynomial discriminant) and to prove in Lean that each algorithm matches the supplied specification for every input, with submissions due November 20, 2026, 23:59 AoE.

The study developed a HALO AI-based placental tissue classifier that uses PLAP and DAPI staining to segment sections into villous core, villous trophoblast (VT), and intervillous space (IVS), validated area and intensity measurement in classified regions with SDC-1 and vimentin staining, and increased the area of staining analyzed to 245 times that of a single field of view on whole slide scanning images.

Using a design-based research framework, this study implemented an AI-driven immediate feedback system with 89 undergraduate judo therapy students during a four-day observation-based clinical placement, where students submitted daily reflective notes and received rubric-based AI scores and feedback within one minute; all 356 system requests were processed successfully, daily note submission rates exceeded 98% and self-assessment completion rates exceeded 95%, and student questionnaires indicated favorable perceptions; as an exploratory external check, three blinded clinical educators independently rated 120 notes from 30 randomly selected students, showing a moderate rank association (Spearman's rho = .495) but limited absolute agreement (ICC = .

This review systematically examines the molecular mechanisms of perineural invasion (PNI) in pancreatic ductal adenocarcinoma, proposes a unified four-stage model spanning mutual chemotaxis between tumors and nerves, adhesion and invasion at the tumor-nerve interface, extracellular matrix remodeling, and neural plasticity alterations, defines the perineural invasion microenvironment as a neuro-immune privileged sanctuary, and summarizes therapeutic strategies targeting the neuro-immune-tumor axis, ongoing clinical trials, and the applications of multi-omics and artificial intelligence in PNI diagnosis, mechanistic discovery, and therapeutic optimization.

This study had five current large language models (GPT 5.3, Claude Sonnet 4.6, Gemini 2.5 Flash, Perplexity, and Grok) reproduce three statistical tasks from a published clinical workflow under identical datasets and standardised prompts—descriptive table generation, Kaplan–Meier survival analysis, and Cox proportional hazards modelling—and compared their outputs with analyses by two experts trained in mathematical statistics on grouping correctness, numerical accuracy, missing-value handling, and quality of generated R code, finding that all five models produced correct descriptive statistics once dataset variables were specified explicitly, that two models failed the initial descriptive benchmark because of variable-name ambiguity but recovered after prompt clarification, that all models

The loaded text contains only the paper title, "RaFT-DM: A Residue-Aware Fusion Transformer With Domain-Wise Memory for Accurate Multi-Label Protein Function Prediction," together with IEEE Xplore navigation, account, copyright, and front-end script template content, with no abstract, methods, experiments, or results, so it can only be confirmed that the work proposes a method named RaFT-DM that combines a residue-aware fusion transformer with domain-wise memory for multi-label protein function prediction, while its specific approach and findings cannot be summarized.

This retrospective study analyzed 1,754 labeled public cystoscopy images to test Direct, Book-based, and Optimized prompts across GPT-5.2, GPT-5, GPT-5-Mini, and GPT-5-Nano for benign-versus-malignant classification, finding that GPT-5 and GPT-5-Mini with the optimized prompt reached accuracies of 86.7% and 89.2%, and that GPT-5 with the optimized prompt achieved 98.1% accuracy, 94.6% specificity, and 99.1% sensitivity in high-confidence triage at 62.0% image coverage, while prompt engineering improved calibration and triage without statistically significant performance gains.

This review surveys the design principles and functional architectures of stimuli-responsive nanoprobes, noting that they can achieve context-responsive activation and signal modulation in response to endogenous tumor cues (acidic pH around 6.5-6.8, glutathione at 2-10 mM, enzymatic overexpression, hypoxia below 2% O2, and redox gradients) and exogenous triggers (near-infrared light at 700-1000 nm, magnetic fields, and ultrasound), yielding 5-20-fold signal amplification, and that they can be integrated with multimodal imaging and artificial intelligence for real-time data interpretation and adaptive diagnostics.

According to this report, Anthropic says researchers may have used its model Claude for nefarious work on viruses and toxins, and this alarming report has divided scientists.

This review, based on a structured literature search of PubMed/MEDLINE, Web of Science and Scopus (2000-2026) supplemented by guidance documents from OECD, FDA, EMA, EFSA and EPA, traces the development of AI in toxicology from rule-based expert systems to deep learning, large language model and multimodal architectures, reports that graph neural networks and multi-task deep learning have shown competitive performance in selected benchmark studies of drug-induced liver injury, hERG cardiotoxicity and Ames mutagenicity, and concludes that inconsistent external validation, limited generalisation of endpoint-specific models and hallucination in large language models leave unresolved regulatory risks, so AI models augment but cannot yet replace experimental toxicology.

This review systematically outlines host-guest composites using metal-organic frameworks (MOFs) as hosts and metal halide perovskites as guests, summarizing three synthetic strategies (ship-in-a-bottle, bottle-around-ship, and one-pot synthesis), noting that nanoconfinement, interfacial passivation, and electronic coupling of MOFs inhibit ion migration, defects, and degradation to improve photoluminescence quantum yield and stability, that emission wavelength and exciton dynamics can be tuned via MOF pore size, ligand functionalization, and nucleation kinetics, and outlining applications in photovoltaics, sensing, information encryption, anti-counterfeiting, and light-emitting devices along with integrating artificial intelligence with theoretical simulation for design optimization.

This technical case report develops and evaluates a system that combines local large language models with rule-based validation against HUGO Gene Nomenclature to extract and structure mutational eligibility criteria from clinical trial records (identifying mutated genes, distinguishing inclusion from exclusion criteria, and assigning them to individual study arms); applied to 4,918 clinical trials it produced structured representations for 1,010 studies, and expert review of 42 trials showed 88.1% of studies correctly annotated with 80% precision at the level of individual eligibility criteria, with failures mainly due to hallucinated genes and misinterpreted abbreviations.

This study developed and validated DiagnosticXchange, an open-source clinical simulation framework in which AI systems diagnose cases by ordering tests, requesting imaging, and performing procedures, with each action mapped to CPT codes capturing cost, time, work relative value units, and invasiveness; using 8 large language models on 216 peer-reviewed cases across 19 specialties (1728 sessions), it found that three systems with near-identical accuracy (93.5%–94.0%) differed significantly in cost (P < .001; 1.75-fold between the most and least expensive) and 2.

This review synthesizes the current literature on explainable and fair artificial intelligence in acute stroke management, noting that AI already performs strongly on tasks such as large vessel occlusion detection, Alberta Stroke Program Early Computed Tomography Score scoring, and functional outcome prediction, while post hoc explainability methods remain approximate and rarely formally tested, fairness evaluation remains uncommon due to limited demographic metadata, regulatory constraints, and the absence of stroke-specific fairness criteria, explainability and fairness remain largely disconnected, and generalizability is affected by dataset partitioning and reporting practices, leading to a call for unified evaluation frameworks that jointly assess explainability, fairness, and generali

In 49 patients with vitreoretinal lymphoma (VRL) confirmed by cytopathology, immunohistochemistry, flow cytometry, and/or MYD88 PCR, host-genome metagenomic deep sequencing (MDS) of intraocular specimens cross-referenced against the Catalogue of Somatic Mutations in Cancer identified eight gene mutations associated with chemotherapeutic resistance in six specimens from four patients, including methotrexate-resistance-associated mutations in four specimens from three patients, and in one patient serial sampling at initial vitrectomy and two subsequent recurrences revealed distinct resistance-associated mutations at each time point; that patient died despite multi-agent therapy including rituximab, consolidation regimens, and lenalidomide, whereas the other three patients remained in long-te

This review systematically surveys the current landscape of artificial intelligence and machine learning in lead optimization, covering advances in graph neural networks, transformer architectures, diffusion models, and chemical foundation models for molecular design and property prediction, and discusses emerging concepts including data-centric AI, uncertainty quantification, trustworthy AI, the AI optimization paradox, and the shift from molecular prediction toward scientific decision-making, concluding that despite increasing industrial adoption, AI remains dependent on high-quality experimental data, model generalizability, and rigorous experimental validation, and that future progress will depend less on increasingly sophisticated algorithms than on trustworthy AI systems that improve

The authors introduce SPHERE, a model-free method that turns sensitive datasets into shareable synthetic twins that AI systems and collaborators can use directly while the original records never leave the local environment; across 33 datasets spanning five scientific domains, SPHERE protects individual privacy against adversarial re-identification attacks while preserving statistical structure (means, variances and correlations reproduced exactly, effect size and P value numerically identical in linear analysis, nonlinear machine-learning utility retained, each twin generated in seconds on a laptop), frontier AI agents on the twin reach the same scientific conclusions as on the original records, analyses reproduce genome- and proteome-wide results at UK Biobank scale and recover landmark f

The study evaluates nine LLM variants across three families (GPT-4.1, GPT-5, Gemini 2.5) on three MoleculeNet datasets (Delaney solubility, Lipophilicity, QM7 atomization energy) using a systematic blinding procedure that iteratively reduces available information, complemented by 0-, 60-, and 1000-shot in-context sample sizes, in order to determine whether molecular property prediction reflects genuine in-context regression or verbatim retrieval of memorized target values, and it adds positive and negative controls for the memorization experiments, structural reference baselines for the multi-shot experiments, and bootstrap confidence intervals for all results; it finds no evidence of verbatim retrieval on these legacy benchmarks and shows that blinding exposes conflicts between pre-traine

Combining a systematic literature review with an experimental evaluation, this study tested ten freely available large language models on five synthetic German clinical notes using standardized prompts, finding that all models substantially increased text length and consistently reduced the density of technical terms and abbreviations, yet no model achieved consistent improvements across all readability indices, with Mistral, ChatGPT, and Copilot showing the highest efficiency in balancing linguistic simplification and text length, suggesting that conventional readability metrics should be extended with domain-specific measures.

This review searched PubMed, Scopus, Web of Science, and Google Scholar for literature up to March 2026 and synthesized the preclinical evidence, safety and regulatory considerations, and clinical development path for intranasal cannabidiol (CBD) delivered via nanostructured lipid carriers (NLCs) in Dravet syndrome (DS) and Lennox-Gastaut syndrome (LGS), noting that intranasal NLC-CBD increased brain CBD levels, enhanced brain targeting, and prolonged central exposure in animal models, while properly designed clinical trials are still needed to establish safety, pharmacokinetics, and efficacy in pediatric DS and LGS patients.

This work introduces an intelligent sequential proximity ligation assay (iseqPLA) read out by spinning disk confocal microscopy and 3D reconstruction to profile endogenous NF-κB protein-protein interactions inside single cells, treating clusters of co-localized puncta as a measure of supercomplex spatial organization, and tracks supercomplex dissociation, p65 nuclear translocation, and negative-feedback engagement across cytokine time courses in NIH-3T3 mouse fibroblasts, cystic fibrosis (CF) patient-derived macrophage co-cultures with IMR-90 human fibroblasts, and an independent set of healthy- and CF-donor monocyte-fibroblast co-cultures, reporting three findings: 3D volumetric quantification reduces the variance in nuclear-to-cytoplasmic ratio measurements relative to 2D projections, th

In a prospective single-center cohort of 143 patients with clinically localized prostate cancer (cT1-2N0M0) undergoing laparoscopic radical prostatectomy, the study measured preoperative serum bone metabolism indicators such as osteocalcin N-terminal mid-fragment and alkaline phosphatase, extracted biparametric MRI radiomics variables including apparent diffusion coefficient mean and K trans mean, and integrated serum and imaging biomarkers with a multivariate logistic regression model; the integrated model reached an area under the ROC curve of 0.984, significantly outperforming individual biomarkers and imaging features (all p < 0.001), showed clinical net benefit with strong calibration (Hosmer-Lemeshow test p = 0.

This review surveys the diverse sources and novel preparation strategies of food-derived antihypertensive peptides, examines their multitarget mechanisms and structure-activity relationships, and summarizes how machine learning supports precise identification and activity prediction, arguing that future work should integrate advanced biotechnologies and intelligent platforms to accelerate the transition from laboratory to clinical application.

This retrospective study submitted standardized summaries of 107 single-center hepatopancreatobiliary (HPB) multidisciplinary team (MDT) cases four times each to four large language models (GPT-4o, GPT-5.2, Gemini 3 Pro, and Claude Sonnet 4.5), finding that response stability differed significantly across models (Gemini 3 Pro lowest discordance, mean 12.8%, Fleiss κ=0.737; GPT-4o highest, mean 30.2%, Fleiss κ=0.430), that concordance with MDT decisions ranged from 48.6% to 72.9% using the initial response and from 66.3% to 74.5% using the modal response, and that class-wise F1 was consistently lower for surgery (0.400–0.520) than for chemotherapy (0.621–0.836), indicating that concordance alone is insufficient for evaluating LLMs as clinical decision support tools.

This work introduces the Virtual Biotech, an organization of AI agents modeled on a drug-development company with agentic divisions spanning target discovery, safety assessment, modality selection, and clinical development, and demonstrates its utility at three drug-development decision points: over 37,000 agents annotated outcomes from 55,984 trials and found that drugs targeting cell-type-specific genes were 48% more likely to reach market with 32% fewer adverse events; it integrated multimodal evidence to propose a therapeutic strategy in lung cancer; and it analyzed a terminated ulcerative colitis trial and inferred potential mechanisms of failure.

This thesis explores how sensor signals combined with AI models can quantify equine locomotion and respiration, finding that a single limb sensor can accurately classify terrain type even at low sampling rates, that a combination of head, withers, and pelvis sensors can discriminate between sound and lame strides with performance varying by front or hind limb lameness, that vertical ground forces are better estimated from upper-body sensors than limb sensors, that dynamic respiratory rate can be computed from downsampled audio signals, and that explainable AI confirms models rely on kinematic landmarks similar to those used by veterinarians, while also presenting a data collection platform for harness racing horses and publishing an audio dataset.

Using up to 30 years of monthly monitoring records from two ecologically contrasting Japanese lakes, the deep-stratified Lake Biwa and the shallow nutrient-rich Lake Kasumigaura, this study benchmarked two time-series foundation models, the Transformer-based Chronos-T5 and the probabilistic Lag-Llama, against 15 statistical (AR, ARIMA, SARIMA, Prophet), machine learning (Random Forest, XGBoost, KNN, SVR), and deep learning (LSTM, CNN, TCN, and SSA-hybrid) approaches, finding that fine-tuned Chronos-T5 ranked first across all six monitoring sites with R2 > 0.80-1.00 for dissolved oxygen, water temperature, pH, and nutrients, while chlorophyll-a remained below R2 0.

Drawing on the author's own experience as a former UX practitioner and AI ethics team member at Microsoft, alongside semi-structured interviews with nine practitioners who had experienced values misalignment in their work on genAI, and analyzing the data through reflexive thematic analysis complemented by 175-word speculative microfictions, this research identifies three themes—practitioners' ambivalence rather than opposition toward genAI, their significant ethical disempowerment due to structural barriers and the absence of ethical discussion in the workplace, and their use of "lean in" and "lean out" tactics to exert agency—and, framed through the metaphor of loops, argues that expecting practitioners to bear ethical accountability for outcomes they cannot meaningfully influence is incr

This report develops CLASS, a Python-based modular large language model pipeline that runs within a secure institutional environment and combines expert-curated concept lists, a task-specific prompt suite, and a schema-constrained output format to extract structured data from clinical notes, and evaluates it exploratorily on a single-center retrospective corpus of pediatric esophageal airway treatment surgery (EATS) operative notes: observed concordance with surgeon adjudication on the 20 longest notes (3,960 note-procedure pairs) was high (F1 0.9967), while CLASS proposed 28 candidate procedure variants or additions, 18 (64.

This study proposes HARU-Net, a hybrid attention residual U-Net trained on a cadaver dataset of human hemimandibles acquired with a high-resolution cone-beam computed tomography (CBCT) protocol for low-dose CBCT denoising; the architecture embeds a hybrid attention transformer block within each skip connection, a residual hybrid attention transformer group at the bottleneck, and residual learning convolutional blocks, and reports the highest peak signal-to-noise ratio of 37.52 dB, the second-highest SSIM of 0.9557, and the lowest GMSD of 0.1084, while maintaining substantially lower computational complexity than transformer-based methods.

This study investigates medical concept normalization of short German clinical expressions to SNOMED CT by comparing a direct GPT-5.4 LLM-only approach with a hybrid approach combining medBERT.de bi-encoder embedding retrieval and RAG-based LLM reranking, finding that the LLM-only baseline achieves Recall@1 of 0.235, Recall@3 of 0.297, and Recall@5 of 0.303, while embedding-based retrieval reaches Recall@1 of 0.681, Recall@3 of 0.783, and Recall@5 of 0.812, and adding RAG reranking further improves Recall@1 to 0.771 with Recall@3 and Recall@5 at 0.809 and 0.812.

The work proposes SurgDepth, a test-time adaptation framework that adapts natural-image-pretrained depth foundation models to surgical endoscopy without any labeled surgical data, combining two self-supervised signals (stereo photometric consistency and flip equivariance, the latter requiring only single images) with selective decoder adaptation and two tuning-free reliability mechanisms (progress-aware anchor regularization and a structural-trust safeguard for poorly-illuminated sequences); the authors report AbsRel reductions of up to 63% with stereo pairs and 42% without, improvement in four of five cross-domain evaluation settings, generalization across seven foundation models, and the lowest AbsRel (0.

This study anonymously surveyed 34 ophthalmologists (median age 49 years, 47% female) from a regional physician network in Germany via online questionnaire and found that most respondents spent 20–39% of their working time on clinical documentation, approximately three quarters expressed general interest in using a reliable and secure ambient scribe system, key perceived benefits were reduced documentation time and increased time for patient care, major barriers were concerns about AI reliability, integration into existing IT systems and data protection, with locally deployed solutions preferred over cloud-based approaches.

This keynote talk examines how hyperlinks can point anywhere and also nowhere, arguing that scholarly content assigned a DOI appears as a hyperlink carrying authority from its independent metadata, and can thereby instantiate “phantom citation objects” describing targets that never existed and never will exist; coupled with the growth of LLM-generated papers and their “rancid hallucinated claims” for cited works, such phantom research objects proliferate throughout the scholarly record, raising questions about the authority claims of DOI hyperlinks and the practical implications for digital preservation.

The authors present MicroClaw, an AI agent that advises on experiment- and system-specific parameters and approaches and collaboratively plans and executes diverse, complex, and reusable imaging workflows across microscope platforms, without requiring expert knowledge.

This update expands the ModelSEED Biochemistry Database to roughly 46,000 compounds, 56,000 reactions and 37,000 metabolic structures, widens thermodynamic handling from two sources to four (group contribution, eQuilibrator 3.0, dGPredictor and experimental values) each kept with its own uncertainty, assigns gold, silver or bronze evidence grades to about 33,000 reactions, and releases reaction directions predicted by an ensemble of large language models alongside a new conflict-resolution pipeline that documents structural choices across sources.

Using the hydrogen evolution reaction (HER) as a prototypical case, this study analyses the limitations of current first-principles approaches, particularly those based on the computational hydrogen electrode (CHE), for predicting single-atom catalyst (SAC) activity, identifying factors such as the sensitivity of reaction thermodynamics to the local atomic environment, the often-unknown experimental structure of SACs, neglected SAC-specific reaction intermediates, solvent effects, catalyst evolution under operating conditions, material instability, and intrinsic density functional theory approximations as sources of discrepancy between theory and experiment, and proposing that integrating these chemical complexities and uncertainties, potentially through artificial intelligence and data-dr

This study combines label-free optical diffraction tomography (ODT) with an interpretable machine-learning framework to extract physically interpretable morphological and biophysical descriptors (sphericity, solidity, eccentricity, dry mass, maximum refractive index) and self-supervised vision transformer (ViT) image representations from three-dimensional refractive index tomograms of red blood cells from synchronized P. falciparum cultures, finding that gametocyte-stage infected RBCs show significantly reduced sphericity and increased eccentricity, with combined features achieving 88.3% accuracy in multiclass classification (normal, ring, gametocyte) and 98.

This study proposes a reference-free evaluation framework that combines embedding-based transcript-summary semantic similarity, stability analysis across repeated generations, and structured human evaluation to benchmark LLM post-processing models for clinical transcripts produced by a locally deployed Whisper system, applying it to 26 German physician-patient conversations in which four LLMs generated 416 summaries, and finding that GPT-OSS-120B and MedGemma-27B achieved the highest similarity scores across most embedding models, that cosine similarities across repeated runs typically exceeded 0.90, and that embedding-based similarity signals only partially agreed with human judgments of summary quality.

This work constructed PlantRegMoD, an AI-powered integrated multi-omics database dedicated to plant regeneration, hosting 20.54 TB of standardized multi-omics data from 147 projects across 32 plant species and 2,593 samples, establishing a unified hierarchical classification system covering five major categories and nine regeneration models, curating 236 regeneration genes and their 28,190 homologs across 58 representative plant species, and containing 196,423 single cells and over 8.81 million epigenetic peaks, equipped with nine online omics tools and a RAG-based intelligent Q&A system.

This pilot study anonymized 35 German doctor's notes from five patients, built one pipeline for medication extraction and mapping and two for diagnoses (one RAG-based and one agentic AI), ran them with three open-weight LLMs on a local GPU-PC, and found that medication name extraction reached an F1 of 0.95 and medication mapping 0.78, while diagnosis coding did not exceed an F1 of 0.12 and broad-category mapping reached 0.18, leading the authors to conclude that LLMs are suitable for medication information extraction for research databases but that current state-of-the-art open-weight models are not accurate enough for a clinical setting where patient treatment would depend on LLM performance.

This work developed RADAR, a generalist vision-language model trained on more than 400,000 contrast-enhanced abdominal CT examinations and 15 million anatomy-wise image-text pairs, learning directly from clinical reports without manual annotation, achieving high diagnostic performance and robust generalization across internal and external evaluations for 18 anatomical structures and 146 imaging findings, and increasing the diagnostic sensitivity of 26 radiologists by ~10% in a reader study.

Following PRISMA-ScR guidelines, this systematic scoping review searched PubMed and Web of Science and identified 72 relevant studies, mapping AI applications in social support for persons with disabilities through the lens of participation, autonomy, and environmental fit across five areas—"Mobility and Navigation Assistance" (n = 20, 41.7%), "Communication and Information Accessibility" (n = 14, 29.2%), "Smart Assistance for Daily Living" (n = 7, 14.6%), "Education and Vocational Empowerment" (n = 5, 8.3%), and "Mental Health Support and Social Inclusion" (n = 3, 6.3%)—and identifying challenges including data privacy (58.3%), inadequate training datasets (25.0%), high implementation costs (25.0%), algorithmic biases (20.8%), and limited real-world evidence of benefits (20.

On September 17, 2026, Anthropic introduced the Life Sciences Verification Program (LSVP), which verifies research credentials, security standards, and ethical research oversight to grant life science teams two types of access—Standard Use and High-risk Use—enabling work such as drug discovery, research biology, clinical development, and manufacturing that is currently blocked in generally-available models, while shifting safeguards from real-time blocking to offline monitoring with a 30-day data retention requirement for LSVP traffic.

Using an on-premise fine-tuned Llama-3.1-70b medication information extraction pipeline, the study automatically extracted medication information from 538 unannotated routine 2012 doctors' letters and compared them with 500 CARDIO:DE letters from 2020/21 (using gold-standard annotations), finding that the DOAC proportion rose from 16.9% to 59.9% while the VKA proportion fell from 37.7% to 9.9%, that the dominant active ingredient within DOACs shifted from rivaroxaban to apixaban, and that manual review showed remaining errors were mainly linked to generic medication mentions and missing medication-reason relations rather than incorrect extraction.

In this cross-sectional comparative study, 49 second-year dental students completed a 45-item multiple-choice final examination in prosthodontic technology under standardized conditions, the same items were submitted to three AI systems (ChatGPT, Gemini, and DeepSeek), and analysis using ANOVA, Cronbach Alpha, and Tukey HSD showed that DeepSeek achieved perfect scores on high-difficulty questions while ChatGPT showed adaptive improvement across attempts, both significantly outperforming the student cohort, whereas Gemini demonstrated lower and consistent accuracy, with significant differences between models (P=0.036, eta-squared=0.891), leading the authors to argue that unsupervised online summative examinations require critical reevaluation of assessment integrity.

In a 401-patient multicentre European prospective pulmonary arterial hypertension (PAH) cohort, adding right ventricular echocardiographic parameters—especially right ventricular-pulmonary arterial (RV-PA) coupling indices—to the ESC/ERS and REVEAL 2.0 risk scores improved the c-index for morbi-mortality prediction from 0.73 to 0.78 and from 0.74 to 0.79, respectively (P<0.01), with machine learning identifying RV-PA coupling as the most influential dimension.

Using 29,530 fundus images from a health screening cohort, this study trained a multi-task model to predict retinal age and evaluated the retinal age gap (RAG) in two sub-cohorts, finding that higher RAG was significantly associated with smoking (ex-smokers beta = +0.46 years; current smokers beta = +0.50 years) and clinical diabetes (+2.52 years), and that RAG was significantly higher in eyes with age-related macular degeneration (+0.60 years) and cataract (+1.86 years) than in normal controls.

This work models iterative generative-model training as a closed-loop stochastic process on the probability simplex and analyzes its dynamics under the Fisher-Rao metric instead of the Euclidean metric, deriving contraction and invariance bounds that remain meaningful as dimension grows and giving a human-to-synthetic data ratio threshold that guarantees convergence to a Fisher-Rao ball, concluding that the effective required human-data ratio is higher than previously implied.

OpenAI and AARP are bringing free, hands-on ChatGPT workshops to 1,000 older adults across 10 U.S. cities to help them build practical AI skills safely.

A report in Cell describes an artificial-intelligence system for ageing biology that introduces large language models trained on ageing data, a suite of 17 benchmark tasks for evaluating these and other LLMs on ageing-related projects, and an interface that brings the models and other ageing research tools together with AI agents; using these benchmarks, the authors compared their specialist LLMs with much larger commercial models from companies including OpenAI and DeepSeek, finding that the ageing-tailored LLMs outperformed the large LLMs in many, though not all, tests.

This news report describes a study published in Nature Neuroscience in which researchers recorded electrical activity directly from the orbitofrontal cortex (OFC) of six people and found that two neighbouring OFC areas are anti-correlated in real time—one becoming more active before participants chose to pursue a reward, the other before they chose to avoid a risk—with signals flickering rapidly between the two regions before settling on a choice in tough decisions.

The study developed RepurAgent, a hierarchical multi-agent AI system in which a supervisor agent and a planning agent coordinate four specialized sub-agents (research, prediction, data, and report) through a human-in-the-loop design with episodic memory and retrieval-augmented generation, and validated it across three scenarios spanning the drug repurposing lifecycle: in Acute Myeloid Leukemia a blinded expert evaluation indicated substantially more novel and mechanistically credible candidates than a vanilla LLM baseline; in a retrospective COVID-19 antiviral screen it prioritized compounds with AUC-ROC up to 0.

A Nature news report says OpenAI announced on 8 September that its AI model solved the Navier–Stokes problem in fluid dynamics and verified the proof using the Lean language, while mathematicians Tristan Buckmaster and Levent Alpöge say they had already been working on the problem with OpenAI and Anthropic tools and Andreas Thom says his discussions about non-sofic groups resembled OpenAI's later approach, prompting an open letter from 25 Fields Medal winners and wider debate about credit and training-data provenance in the AI era.