Nature Medicine, Published online: 12 January 2026; doi:10.1038/s41591-025-04126-3

Including data from 1,047 patients across 19 inflammatory diseases, a new atlas presents a comprehensive model of inflammation in circulating immune cells.

arXiv:2601.05148v1 Announce Type: cross Abstract: Pathology foundation models substantially advanced the possibilities in computational pathology -- yet tradeoffs in terms of performance, robustness, and computational requirements remained, which limited their clinical deployment. In this report, we present Atlas 2, Atlas 2-B, and Atlas 2-S, three pathology vision foundation models which bridge these shortcomings by showing state-of-the-art performance in prediction performance, robustness, and resource efficiency in a comprehensive evaluation across eighty public benchmarks. Our models were trained on the largest pathology foundation model dataset to date comprising 5.5 million histopathology whole slide images, collected from three medical institutions Charit\'e - Universt\"atsmedizin Berlin, LMU Munich, and Mayo Clinic.

Nature, Published online: 10 December 2025; doi:10.1038/d41586-025-04037-w

Analysis of more than one million people shows that mental-health disorders fall into five clusters, each of them linked to a specific set of genetic variants.

arXiv:2510.18212v3 Announce Type: replace Abstract: The lack of a concrete definition for Artificial General Intelligence (AGI) obscures the gap between today's specialized AI and human-level cognition. This paper introduces a quantifiable framework to address this, defining AGI as matching the cognitive versatility and proficiency of a well-educated adult. To operationalize this, we ground our methodology in Cattell-Horn-Carroll theory, the most empirically validated model of human cognition. The framework dissects general intelligence into ten core cognitive domains-including reasoning, memory, and perception-and adapts established human psychometric batteries to evaluate AI systems. Application of this framework reveals a highly "jagged" cognitive profile in contemporary models. While proficient in knowledge-intensive domains, current AI systems have critical deficits in foundational cognitive machinery, particularly long-term memory storage. The resulting AGI scores (e.g., GPT-4 at 27%, GPT-5 at 57%) concretely quantify both rapid progress and the substantial gap remaining before AGI.

arXiv:2511.11733v1 Announce Type: cross Abstract: Speculative decoding accelerates large language model (LLM) inference by using a lightweight draft model to propose tokens that are later verified by a stronger target model. While effective in centralized systems, its behavior in decentralized settings, where network latency often dominates compute, remains under-characterized. We present Decentralized Speculative Decoding (DSD), a plug-and-play framework for decentralized inference that turns communication delay into useful computation by verifying multiple candidate tokens in parallel across distributed nodes. We further introduce an adaptive speculative verification strategy that adjusts acceptance thresholds by token-level semantic importance, delivering an additional 15% to 20% end-to-end speedup without retraining. In theory, DSD reduces cross-node communication cost by approximately (N-1)t1(k-1)/k, where t1 is per-link latency and k is the average number of tokens accepted per round. In practice, DSD achieves up to 2.56x speedup on HumanEval and 2.59x on GSM8K, surpassing the Eagle3 baseline while preserving accuracy. These results show that adapting speculative decoding for decentralized execution provides a system-level optimization that converts network stalls into throughput, enabling faster distributed LLM inference with no model retraining or architectural changes.

Nature Biotechnology, Published online: 07 November 2025; doi:10.1038/s41587-025-02941-0

Publisher Correction: Deep-learning-based virtual screening of antibacterial compounds

arXiv:2511.02531v1 Announce Type: cross Abstract: Healthcare artificial intelligence systems routinely fail when deployed across institutions, with documented performance drops and perpetuation of discriminatory patterns embedded in historical data. This brittleness stems, in part, from learning statistical associations rather than causal mechanisms. Causal graph neural networks address this triple crisis of distribution shift, discrimination, and inscrutability by combining graph-based representations of biomedical data with causal inference principles to learn invariant mechanisms rather than spurious correlations. This Review examines methodological foundations spanning structural causal models, disentangled causal representation learning, and techniques for interventional prediction and counterfactual reasoning on graphs. We analyse applications demonstrating clinical value across psychiatric diagnosis through brain network analysis, cancer subtyping via multi-omics causal integration, continuous physiological monitoring with mechanistic interpretation, and drug recommendation correcting prescription bias. These advances establish foundations for patient-specific Causal Digital Twins, enabling in silico clinical experimentation, with integration of large language models for hypothesis generation and causal graph neural networks for mechanistic validation. Substantial barriers remain, including computational requirements precluding real-time deployment, validation challenges demanding multi-modal evidence triangulation beyond cross-validation, and risks of causal-washing where methods employ causal terminology without rigorous evidentiary support. We propose tiered frameworks distinguishing causally-inspired architectures from causally-validated discoveries and identify critical research priorities making causal rather than purely associational claims.

arXiv:2510.22620v1 Announce Type: cross Abstract: AI agents powered by large language models (LLMs) are being deployed at scale, yet we lack a systematic understanding of how the choice of backbone LLM affects agent security. The non-deterministic sequential nature of AI agents complicates security modeling, while the integration of traditional software with AI components entangles novel LLM vulnerabilities with conventional security risks. Existing frameworks only partially address these challenges as they either capture specific vulnerabilities only or require modeling of complete agents. To address these limitations, we introduce threat snapshots: a framework that isolates specific states in an agent's execution flow where LLM vulnerabilities manifest, enabling the systematic identification and categorization of security risks that propagate from the LLM to the agent level. We apply this framework to construct the $\operatorname{b}^3$ benchmark, a security benchmark based on 194331 unique crowdsourced adversarial attacks. We then evaluate 31 popular LLMs with it, revealing, among other insights, that enhanced reasoning capabilities improve security, while model size does not correlate with security. We release our benchmark, dataset, and evaluation code to facilitate widespread adoption by LLM providers and practitioners, offering guidance for agent developers and incentivizing model developers to prioritize backbone security improvements.

npj Digital Medicine, Published online: 27 October 2025; doi:10.1038/s41746-025-01996-2

Benchmarking large language models for personalized, biomarker-based health intervention recommendations

arXiv:2510.18212v2 Announce Type: replace Abstract: The lack of a concrete definition for Artificial General Intelligence (AGI) obscures the gap between today's specialized AI and human-level cognition. This paper introduces a quantifiable framework to address this, defining AGI as matching the cognitive versatility and proficiency of a well-educated adult. To operationalize this, we ground our methodology in Cattell-Horn-Carroll theory, the most empirically validated model of human cognition. The framework dissects general intelligence into ten core cognitive domains-including reasoning, memory, and perception-and adapts established human psychometric batteries to evaluate AI systems. Application of this framework reveals a highly "jagged" cognitive profile in contemporary models. While proficient in knowledge-intensive domains, current AI systems have critical deficits in foundational cognitive machinery, particularly long-term memory storage. The resulting AGI scores (e.g., GPT-4 at 27%, GPT-5 at 57%) concretely quantify both rapid progress and the substantial gap remaining before AGI.

New AI browsers from OpenAI and Perplexity promise to increase user productivity, but they also come with increased security risks.

npj Digital Medicine, Published online: 21 October 2025; doi:10.1038/s41746-025-02082-3

Although still at a nascent state, quantum computing promises advances in healthcare, from drug discovery to personalised treatments. But it also threatens current cryptographic systems that protect medical data and infrastructure. The concept of “Q-Day” highlights risks such as “harvest now, decrypt later” attacks, with particular concerns for medical devices and sensitive applications in fields like femtech. Preparing for this future requires the rapid adoption of post-quantum cryptography, the coordination of time-phased and scalable “technology rollout” strategies, and revised regulatory frameworks to safeguard patient safety, privacy, and trust.

The White House's David Sacks and OpenAI's Jason Kwon caused a stir online this week for their comments about groups promoting AI safety.

Science, Ahead of Print.

A new test from OpenAI aims to understand how close AI is to outperforming humans at economically valuable work.

Background: Large language models (LLMs) hold promise for supporting clinical tasks, particularly in data-driven and technical disciplines such as radiation oncology. While prior evaluation studies have focused on examination-style settings for evaluating LLMs, their performance in real-life clinical scenarios remains unclear. In the future, LLMs might be used as general AI assistants to answer questions arising in clinical practice. It is unclear how well a modern LLM, locally executed within the infrastructure of a hospital, would answer such questions compared with clinical experts. Objective: This study aimed to assess the performance of a locally deployed, state-of-the-art medical LLM in answering real-world clinical questions in radiation oncology compared with clinical experts. The aim was to evaluate the overall quality of answers, as well as the potential harmfulness of the answers if used for clinical decision-making. Methods: Physicians from 10 departments of European hospitals collected questions arising in the clinical practice of radiation oncology. Fifty of these questions were answered by 3 senior radiation oncology experts with at least 10 years of work experience, as well as the LLM Llama3-OpenBioLLM-70B (Ankit Pal and Malaikannan Sankarasubbu). In a blinded review, physicians rated the overall answer quality on a 5-point Likert scale (quality), assessed whether an answer might be potentially harmful if used for clinical decision-making (harmfulness), and determined if responses were from an expert or the LLM (recognizability). Comparisons between clinical experts and LLMs were then made for quality, harmfulness, and recognizability. Results: There were no significant differences between the quality of the answers between LLM and clinical experts (mean scores of 3.38 vs 3.63; median 4.00, IQR 3.00-4.00 vs median 3.67, IQR 3.33-4.00; P=.26; Wilcoxon signed rank test). The answers were deemed potentially harmful in 13% of cases for the clinical experts compared with 16% of cases for the LLM (P=.63; Fisher exact test). Physicians correctly identified whether an answer was given by a clinical expert or an LLM in 78% and 72% of cases, respectively. Conclusions: A state-of-the-art medical LLM can answer real-life questions from the clinical practice of radiation oncology similarly well as clinical experts regarding overall quality and potential harmfulness. Such LLMs can already be deployed within the local hospital environment at an affordable cost. While LLMs may not yet be ready for clinical implementation as general AI assistants, the technology continues to improve at a rapid pace. Evaluation studies based on real-life situations are important to better understand the weaknesses and limitations of LLMs in clinical practice. Such studies are also crucial to define when the technology is ready for clinical implementation. Furthermore, education for health care professionals on generative AI is needed to ensure responsible clinical implementation of this transforming technology.

OpenAI is continuing to build out massive AI data centers to train and serve increasingly powerful AI models.

Background: Artificial intelligence (AI) and Large Language models (LLMs) are increasing in sophistication and are being integrated into many disciplines. The potential for LLMs to augment clinical decisions is an evolving area of research. Objective: This study compared the responses of over 1000 kidney specialist physicians (nephrologists) to outputs of commonly used LLMs using a questionnaire determining when a kidney biopsy should be performed. Methods: This research group completed a large online questionnaire for nephrologists to determine when a kidney biopsy should be performed. The questionnaire was co-designed with patient participation, refined through multiple iterations, then piloted locally before international dissemination. It was the largest international study in the field and demonstrated variation between human clinicians in biopsy propensity relating to human factors such as sex and age, as well as systemic factors such as country, job seniority and technical proficiency. The same questions were put to both human doctors and LLMs in an identical order in a single session. Eight commonly used LLMs were interrogated: Chat GPT 3.5, Mistral Hugging Face, Perplexity, Microsoft Co-pilot, Llama 2, GPT 4.0, MedLM and Claude 3. The most common response given by clinicians (human mode) to each question was taken as the baseline for comparison. Questionnaire responses to the indications and contraindications for biopsy generated a score (0-44) reflecting biopsy propensity, in which a higher score was used as a surrogate marker for an increased tolerance of potential associated risks. Results: The ability of LLMs to reproduce human expert consensus varied widely with some models demonstrating a balanced approach to risk in a similar manner to humans, whilst other models reported outputs at either end of the spectrum for risk tolerance. In terms of agreement with the human mode, Chat GPT 3.5 and GPT 4.0 (Open AI) had the highest levels of alignment, with the human mode selected in 6/11 questions. The total biopsy propensity score generated from the human mode was 23/44. Both Open AI models produced similar propensity scores between 22 and 24, however Llama 2 and MS Co-pilot also reported scores within this range, but with poorer response alignment to the human mode at only 2/11 questions. The most risk averse model in this study was MedLM with a propensity score of 11 and the least risk averse model was Claude 3 with a score of 34. Conclusions: LLM outputs demonstrated a modest ability to replicate human clinical decision making in this study, however the performance varied widely between LLM models. Questions with more uniform human responses produced LLM outputs with greater alignment, whereas in questions with low levels of human consensus there was poor output alignment. This may limit the practical use of LLMs in real world clinical practice.

Single-cell transcriptomics analysis and multimodal profiling (STAMP) by imaging enables single-cell analysis of cells in suspension without the need for sequencing. The markedly reduced costs and flexible experimental designs support the profiling of millions of cells or the large-scale multiplexing of conditions, perturbations, and sample types.