Much of the human genome’s non-protein-coding fraction acts directly through RNA, yet the structural and functional roles encoded in these sequences remain poorly understood. Applying deep learning is hindered by scarce RNA structural data and it remains unclear what biological constraints such models can recover directly from the abundant RNA sequences alone. Here, to address these challenges, we developed NucleicBERT, a self-supervised masked-language model that learns contextual representations from single sequences without evolutionary information. Explainable artificial intelligence analyses show that the model organizes RNA sequences in latent space and encodes structural properties indicating that biologically meaningful constraints are learned from sequence correlations alone. When fine-tuned for downstream structural and functional tasks, NucleicBERT requires only single sequences while matching or exceeding current RNA prediction models. This alignment-free framework addresses the scarcity of annotated 3D RNA data while providing a rapid, computational complement to experimental techniques. By bridging abundant unlabelled sequence data with scarce structural annotations, NucleicBERT advances RNA structure prediction and informs how large language models encode biological information. RNA structure and function are hard to infer because annotations are scarce, despite abundant sequence data. Upadhyay et al. trained a self-supervised model on large-scale RNA data that derives biologically meaningful patterns from sequence correlations.
While classical neural operators have transformed the solving of partial differential equations, developing efficient quantum counterparts remains challenging due to processing overheads and theoretical gaps. Specifically, existing quantum machine learning paradigms often demand prohibitive qubit scaling or deep circuits, rendering them impractical for near-term hardware. Here, to bridge this gap, we propose QuanONet, a quantum neural operator tailored for the noisy intermediate-scale quantum era. Theoretically, we extend the universal approximation theorem to the quantum domain for continuous nonlinear operators. Departing from conventional views relying on exponential Hilbert spaces, we prove that the architecture’s density matrix implicitly constructs a quadratic feature frame. We establish that, for operators, this implicit frame yields an $${\mathcal{O}}({p}^{2})$$ expressivity bound, circumventing the $${\mathcal{O}}(p)$$ linear capacity limits of matched classical models. To overcome the depth cost of spectrum alignment, we introduce a trainable-frequency (TF-QuanONet) strategy that adaptively spaces base frequencies to capture relatively high frequencies without parameter inflation. Extensive benchmark experiments demonstrate that TF-QuanONet notably outperforms quantum baselines and achieves competitive accuracy against classical frameworks under strictly matched-parameter conditions. Furthermore, in high-dimensional scaling regimes (p → 256), the architecture exhibits superior optimization robustness, consistently converging to the intrinsic error floor where classical baselines suffer from high variance. The physical deployment on IBM quantum processors validates its functional resilience on near-term hardware as a qualitative proof of concept. Wang et al. introduce a hardware-efficient quantum neural operator that overcomes classical linear capacity limits. Using an implicit quadratic frame, it offers accelerated expressivity for solving differential equations in the noisy intermediate-scale quantum era.
Representational alignment can reveal similarities between human brain activity and language models. Work now demonstrates that it can also guide learning, improving the reliability of artificial reasoning.
Clinical prediction models often underperform in deployment populations that differ from historical training cohorts because of covariate shifts, but outcome-based updating is infeasible when deployment outcomes are unavailable. This challenge is further amplified when the target training cohort has limited samples or sparse events. Here we propose CoxRTL, a transfer learning framework with deployment-covariate recalibration for survival prediction under covariate shift. CoxRTL borrows information from external cohorts to improve estimation, uses transferability screening and debiasing to mitigate external effect heterogeneity, and embeds density-ratio weighting into transfer estimation, yielding a distribution-adaptive weighted partial likelihood that aligns historical training cohorts with the deployment-covariate distribution. We evaluated CoxRTL in simulations and two cohorts. Simulations showed improved prediction under training–deployment covariate shift. Using early National Health and Nutrition Examination Survey data, we developed prognostic models for 20 chronic diseases in older adults and evaluated their performance in later survey cohorts. CoxRTL achieved higher discrimination than models trained only in older adults (median C-index improvement, 0.056; range, 0.016–0.126) or in the full population (0.042; 0.013–0.107). It also improved calibration, clinical utility and risk stratification, while identifying variables with greater predictive contribution across diseases. Independent replication in the Shanghai Suburban Adult Cohort and Biobank confirmed robustness under geographic and temporal shifts. These findings support CoxRTL as a feasible strategy for developing prediction models for prespecified deployment populations when target training data are limited. Pan et al. present CoxRTL, a recalibrated transfer learning strategy that leverages external cohorts to improve survival prediction under covariate shift when target training data are limited and deployment outcomes are unavailable.
Artists are often recognizable through collections of distinctive patterns (‘fingerprints’) in their work. Identifying such traits has important applications in authorship attribution, education, cultural heritage research and historical analysis. Here we focus on music, a domain with a rich tradition of theoretical and mathematical analysis. We train a variety of supervised learning models to identify 20 iconic jazz musicians from a curated dataset of 84 h of recordings. In particular, we introduce a multi-input architecture that represents four musical domains separately: melody, harmony, rhythm and dynamics. This design allows us to accurately identify individual performers (our best model obtains 94% accuracy across 20 classes) and to examine which musical elements most strongly distinguish between individual artists. We release open-source implementations of our models and an accompanying web application for exploring our results. Cheston et al. develop a machine learning pipeline that identifies 20 iconic jazz pianists from audio recordings with up to 94% accuracy, revealing how melody, harmony, rhythm and dynamics shape each performer’s individual musical fingerprint.
Computer audition aims to enable machines to understand, process and generate human-audible signals, including speech, environmental sound and music. The field is now evolving beyond traditional paradigms to fully leverage the capabilities of foundation models, towards more comprehensive understanding, more natural generation and more human-like interaction. Audio, as a modality rich in semantic, emotional and contextual cues, has a vital role in achieving naturalistic and embodied machine intelligence. This Review provides a comprehensive review of recent progress in integrating audio into large language models, with a focus on four key areas: audio comprehension, audio generation, speech-based interaction and audio–visual understanding. We analyse how large language models are reshaping audio perception and reasoning, enabling systems to understand sound at a deeper semantic level, generate expressive audio outputs and engage in human-like spoken interaction. Furthermore, we explore how the fusion of audio and visual modalities enhances situational awareness and cross-modal reasoning, pushing the boundaries of multimodal intelligence. This Review not only synthesizes existing research but also identifies critical challenges and future directions for building general auditory intelligence capable of perceiving, understanding and interacting through sound as naturally as humans do. Wang et al. summarize advances in machine listening and speaking, speech-based interaction, and audio–visual understanding.
Knowledge editing has emerged as a promising approach that leverages understanding of a model’s inner knowledge mechanisms to enable precise knowledge updates and behaviour control without costly retraining. It is particularly appealing for enabling continuous knowledge adaptation, a capability essential for building truly intelligent, self-evolving AI systems. However, current methods treat large language models as modular knowledge stores where facts can be edited independently, ignoring the fact that knowledge forms an interconnected system in which elements depend on each other. As large language models increasingly exhibit sophisticated reasoning abilities, such as spanning multistep deduction and causal inference, the need for reasoning-consistent knowledge updates becomes critical. In this Perspective we explore some limitations of existing knowledge editing techniques and argue that effective knowledge editing must account for the intricate nature of knowledge representation. We outline three promising research directions: (1) addressing knowledge interdependence through deductive closure circuit editing; (2) integrating model beliefs and confidence into the editing process; and (3) enabling contextualized updates for complex, interdependent knowledge forms. Together, these directions suggest a pathway towards more principled knowledge editing methods capable of supporting the next generation of adaptive, reasoning-driven AI systems. Chen et al. explore limitations of current knowledge editing techniques in large language models and propose three promising research directions that respect the complexity of knowledge representation in a real-world setting.
A long-standing goal in robotics is to create general-purpose agents that can learn foundational skills applicable across diverse hardware and environments. In manipulation, this is hindered by a fundamental mismatch: deep learning policies are data-efficient but hardware-specific, whereas analytical planners are hardware-agnostic but fail under perceptual uncertainty. Here we introduce SpaHybGen, a framework that unifies their strengths through learned, universal contact representations. We first train a neural network to infer spatial contact features—a hardware-agnostic representation of potential grasp points—directly from noisy depth observations. These features then guide a differentiable optimizer that computes stable grasps for articulated hand models. This hybrid design enables zero-shot generalization: our system, trained once, successfully empowered seven distinct robotic hands (from two to five fingers) without any hardware-specific retraining, achieving grasping success rates of 94.3%–98.0% in semi-cluttered scenes. It further enabled dynamic grasping at 20 Hz in dense clutter and multi-hand coordination for complex tasks. By decoupling perception from action through a shared contact interface, we provide a pathway towards reusable and adaptable manipulation intelligence, a key step for general-purpose robotics. Our code and models are open-sourced to support this vision. Wang et al. design efficient robot–environment interaction representations that achieve generalization across articulated robotic hand models and task adaptability in diverse cluttered grasping scenarios, suggesting a path towards general-purpose robotics.
A new deep learning method enables molecular dynamics simulations over longer time scales while still achieving accurate physical property prediction.
Generative machine learning methods have led to progress in crystal discovery, but cannot fully explore the space of material candidates that are both novel and useful. A reinforcement learning-based method steers candidate generation to these areas, enabling the design of novel functional materials.
Agents, language model-based systems that can reason, plan and act with tools to accomplish tasks, are widely deployed, yet it remains unclear when multi-agent coordination outperforms a strong single agent. Here we conduct a controlled experiment that holds task prompts, tools and compute budgets constant while varying only coordination structure and model capability. Across 260 configurations spanning six benchmarks, five architectures and three LLM families, we derive a predictive model using empirical coordination metrics. Across benchmarks, single-agent baseline performance emerges as the most robust predictor of whether coordination improves or decreases performance. In particular, we identify an empirical capability-saturation threshold beyond which additional agents are unlikely to improve performance. This threshold correctly predicts the effect of multi-agent coordination on performance in 94% of validation configurations on SWE-bench Verified and Terminal-Bench. We therefore interpret this threshold as a practical selection rule rather than a universal scaling principle. A second effect, baseline-scaled error amplification, survives cluster-robust inference (Probust = 0.030) and supports the failure-mode taxonomy. The fitted model achieves cross-validated R2 = 0.373 (0.413 with a task-grounded capability metric) and selects the best architecture in 87% of held-out configurations. These results provide a quantitative framework for within-domain architecture selection and for estimating when multi-agent coordination is likely to improve performance or add overhead. A controlled study of large language model agents across 260 configurations shows when multi-agent collaboration helps or hurts performance, and introduces a predictive model that selects the best architecture in 87% of held-out within-domain configurations.
Artificial intelligence systems are becoming more intelligent, but at a very high cost in terms of energy consumption and training requirements. By contrast, our brains only require 20 W of energy, they learn online and they can instantly adjust to changing contingencies. This begs the question what data structures, algorithms and learning methods enable brains to achieve that, and whether these can be ported into artificial devices. We are addressing this question for a core feature of intelligence: the capacity to plan and solve problems, including new problems that involve states that were never encountered before. Here we examine three tools that brains are likely to use for achieving that: cognitive maps, stochastic computing and compositional coding. We integrate these tools into a transparent neural network model, and demonstrate its power for flexible planning and problem-solving. Importantly, this approach is suitable for implementation by in-memory computing and other energy-efficient neuromorphic hardware. In particular, it only requires self-supervised local synaptic plasticity that is suited for on-chip learning. Hence, a core feature of brain intelligence-the capacity to generate solutions to problems that were never encountered before-does not require deep neural networks or large language models, and can be implemented in energy-efficient edge devices.
Understanding how information is represented in neural networks is a fundamental challenge in both neuroscience and artificial intelligence. Despite their nonlinear computations, ample evidence suggests that neural networks encode features in superposition, meaning that systems linearly represent more concepts than they have neurons. This observation opens the door to extracting interpretable representations from otherwise opaque networks, but a principled account of why superposition arises and how it can be exploited has been lacking. Here we synthesize insights from identifiability theory, compressed sensing and quantitative interpretability research to propose a unified perspective on this phenomenon. Our synthesis yields a three-step framework: identifiability theory establishes that neural networks trained for classification recover latent features up to linear mixing, compressed sensing provides guarantees for disentangling these features via sparse coding, and interpretability metrics grounded in behavioural tasks assess whether the extracted features align with human-interpretable concepts. By bridging theoretical neuroscience, representation learning and interpretability research, our framework connects longstanding questions about neural coding in biological systems with modern efforts in artificial intelligence transparency, and highlights open problems at their intersection. Kindt et al. present a unifying framework for superposition in neural networks. Their three-step approach clarifies how latent features can be identified, disentangled and assessed.
Deep learning has achieved remarkable success in computer vision and natural language processing, where tasks are commonly formulated as mappings between finite-dimensional representations. Many scientific problems, however, including those governed by partial differential equations, are naturally posed on infinite-dimensional function spaces. This mismatch has limited conventional neural networks from achieving comparable success in scientific applications. Here we identify and distil key principles for constructing practical neural architectures for mappings between function spaces. Neural operators provide a principled extension of neural networks to such settings, offering a path towards bringing deep learning's transformative impact to science. Because deep learning's success has relied heavily on architectural refinements, extending these advances to neural operators allows operator learning to benefit from refined designs. Guided by the principles we outline, we propose a recipe for converting popular neural architectures into neural operators with minimal modifications. We also discuss practical steps for making these models effective. This perspective offers a systematic bridge between finite-dimensional network design and operator learning for scientific applications.
Compared with generic artificial intelligence agents, deep research agents perform longer-horizon reasoning and deeper literature exploration to investigate complex questions. Here we present DeepEvidence, a deep research agent for evidence exploration and synthesis across heterogeneous biomedical knowledge sources. DeepEvidence advances deep research through coordinated multi-agent collaboration combining breadth-first and depth-first research strategies to search, explore and aggregate evidence from multiple biomedical knowledge bases and literature. It also incrementally constructs an evidence graph of key entities and observations to support transparent tracking, attribution and validation of the research process. DeepEvidence substantially outperforms generic artificial intelligence agents across four open benchmarks. We further establish seven benchmark tasks spanning major stages of biomedical discovery, including drug discovery, preclinical experimentation, clinical trial development and evidence-based medicine. DeepEvidence demonstrates substantial improvements in systematic evidence exploration and synthesis. These results highlight the potential of deep research agents to accelerate biomedical discovery and translational research.
Executing experimental tasks in both normal research laboratories and large-scale scientific facilities often requires extensive human supervision and remains a key challenge on the path to fully autonomous, artificial intelligence (AI)-driven science. Here we demonstrate a large language model-driven agent that autonomously performs X-ray sample alignment on a synchrotron beamline by planning actions, executing instrumental commands, interpreting observations and iterating towards experimental goals. Based on existing large language models with structured tool-use via the model context protocol, our AI X-ray scientist was guided and tested using an in-house-built virtual experimental setup that mirrors a six-circle diffractometer at an operational synchrotron beamline. The agentic workflow developed in the virtual environment was directly deployed on a real beamline, where it correctly identified reference reflections and determined the orientation matrix, an essential first step in any type of single-crystal scattering experiment. Our AI X-ray scientist responded effectively to unexpected experimental conditions, demonstrating adaptive problem-solving and readiness for addressing practical experimental situations. Our study provides a step towards autonomous operation across diverse experimental environments at large-scale scattering facilities.
Rising pathogen drug resistance makes next-generation antimicrobial peptides a global priority. Generative AI accelerates discovery by rapidly proposing new peptides with high therapeutic potential. The key question is no longer whether broad data-driven exploration is possible, but whether it can refine biologically complex activity scaffolds.
Soft hand exoskeletons have emerged as promising assistive devices for individuals with impaired hand function. However, most existing systems provide limited dexterity and primarily target users with moderate hand ability, leaving individuals with severe hand paralysis without effective solutions for reliable grasping of diverse objects. Here we report the translational development of a lightweight, textile-based soft robotic exoskeleton glove with wrist dorsiflexion and an active opposable and abductable thumb, designed to restore hand function in a patient with severe right-hand impairment due to amyotrophic lateral sclerosis. We followed a co-creation approach, enhancing dexterity by increasing hand articulations based on patient needs. Furthermore, to enhance the patient's sense of control, a non-invasive surface electromyography-based grasp predictor (97% sensitivity) was combined with motion data and machine learning-based error correction to compensate for weak, noisy muscle signals, compared with healthy controls (n = 15). The exoskeleton enabled the patient to grasp objects, achieve a Box-and-Blocks Test score of 5 and perform meaningful tasks, including feeding himself. We further validated the exoskeleton in patients with stroke (n = 6). While exoskeleton assistance on average reduced Action Research Arm Test scores of moderately impaired patients by 9, severely impaired patients scored 17 points higher when using the exoskeleton. These results indicate that the dexterous soft hand exoskeleton is particularly effective for individuals with severe to near-complete hand paralysis, while its utility for patients with moderate residual function is limited and task dependent.
