Nature Machine Intelligence最新文献

{"title":"Implicit-bias-like patterns in reasoning models","authors":"Messi H. J. Lee, Calvin K. Lai","doi":"10.1038/s42256-026-01300-1","DOIUrl":"10.1038/s42256-026-01300-1","url":null,"abstract":"Implicit biases refer to automatic mental processes that shape perceptions, judgements and behaviours. While previous research on bias in large language models (LLMs) has focused primarily on outputs, we introduce the reasoning-model implicit association test (RM-IAT) to study bias-like processing differences in reasoning models (that is, LLMs that generate explicit step-by-step reasoning before producing a response). By measuring reasoning-token counts as an index of computational effort, the RM-IAT captures processing efficiency differences analogous to response latency differences in the human IAT. Across four models (o3-mini, DeepSeek-R1, gpt-oss-20b and Qwen3-8B), we find consistent evidence that association-incompatible tasks require greater computational effort than association-compatible tasks. Claude 3.7 Sonnet exhibited reversed patterns, which were linked using thematic analysis to its unique internal focus on reasoning about bias and stereotypes. We also found evidence for convergent validity with model outputs. RM-IAT effects predicted biases in two tasks known to capture LLM biases in word association and decision-making. Together, these findings demonstrate that the RM-IAT captures meaningful variation in how reasoning models process stereotypical information, and that this variation predicts downstream model behaviour. Lee and Lai study bias-like processing differences in large language reasoning models and find that, for most models, processing stereotypical information takes less computational effort than processing counter-stereotypical information.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 9","pages":"1457-1465"},"PeriodicalIF":29.8,"publicationDate":"2026-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148899595","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Large language models as uncertainty-calibrated optimizers for experimental discovery","authors":"Bojana Ranković, Ryan-Rhys Griffiths, Philippe Schwaller","doi":"10.1038/s42256-026-01283-z","DOIUrl":"10.1038/s42256-026-01283-z","url":null,"abstract":"From reaction optimization to molecular design, experimental discovery poses the same expensive question: which candidate to test next under time and resource constraints. Bayesian optimization provides principled answers but depends on domain expertise that rarely transfers. Large language models (LLMs) contain rich scientific knowledge but lack the calibrated uncertainty estimates crucial for high-stakes decisions. Here we show how training language models through Bayesian objectives enables their use as reliable optimizers guided by natural language. Our approach, GOLLuM (Gaussian process Optimized LLMs), teaches LLMs from experimental outcomes under uncertainty, transforming their overconfidence from a fundamental flaw into a precise learning signal. This signal reshapes the LLM embeddings so that experiments with similar outcomes cluster together, revealing structure in the design space. Starting from only ten low-performing experiments, GOLLuM generalizes across 23 tasks in organic synthesis, materials science, process chemistry and molecular design, ranking first on average among all competing methods. It matches traditional Bayesian optimization with over 40% fewer experiments and nearly doubles the discovery of high-performing Buchwald–Hartwig reactions over expert quantum-chemical descriptors and state-of-the-art LLMs (43% versus 24–25%). More broadly, GOLLuM points to a different paradigm for specializing foundation models: not through more data but through richer, uncertainty-guided information. Although language models can be helpful in molecular design, they are not typically calibrated for uncertainty. Rankovic and colleagues present a method to train language models while taking into account the uncertainty of the data.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 9","pages":"1466-1477"},"PeriodicalIF":29.8,"publicationDate":"2026-08-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-026-01283-z.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148853491","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Enhancing reproducibility in hybrid Earth system models","authors":"Min Chen, Zhiyi Zhu, Thorsten Wagener, Niklas Boers, R. Dietmar Müller, Josef Strobl, Gustau Camps-Valls, Michael Batty, Anthony J. Jakeman, Olaf Kolditz, Stefano Nativi, Maria Antonia Brovelli, Felix Creutzig, Pankaj Kumar, Paul Whitehead, C. Michael Barton, Dichen Liu, Peilong Ma, Zaiyang Ma, Fengyuan Zhang, Bo Zhang, Peng Hou, Guonian Lü","doi":"10.1038/s42256-026-01299-5","DOIUrl":"10.1038/s42256-026-01299-5","url":null,"abstract":"The integration of artificial intelligence into Earth system models (ESMs) has revolutionized the simulation and prediction of complex environmental dynamics. However, this shift introduces substantial challenges for reproducibility, a cornerstone of scientific progress. In particular, artificial intelligence-infused hybrid ESMs face amplified issues of numerical instability, procedural opacity and asymmetric access to computational resources. If left unaddressed, these challenges risk turning hybrid ESMs into opaque and weakly verifiable systems, reducing model traceability, weakening cumulative knowledge building and narrowing the evidential basis for climate risk assessment and policy guidance. This Perspective argues that reproducibility should be reframed to reflect the epistemological and operational realities of hybrid ESMs. We propose an integrated roadmap that couples a theory of reproducibility assessment with practical pathways for implementation in modelling practices. Within this context, we introduce Reproducibility in hybrid Earth system models (RHEM) as a reference guideline for governing transparent, trustworthy, and reproducible hybrid ESMs. Building on this foundation, the pathways operationalize the framework’s criteria into actionable measures that embed transparency and trustworthiness in the modelling process itself. By linking conceptual structure with operational guidance, reproducibility is repositioned from a post hoc requirement to a structural property of hybrid ESMs and established as a foundational principle for Earth system science in the artificial intelligence era. AI integration in Earth system models enhances prediction and modelling capabilities but also amplifies challenges for reproducibility. This Perspective introduces a framework for assessing reproducibility and provides practical ways to strengthen reproducibility in hybrid Earth system models.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 9","pages":"1358-1370"},"PeriodicalIF":29.8,"publicationDate":"2026-08-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148987148","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Life-inspired interoceptive artificial intelligence for autonomous and adaptive agents","authors":"Sungwoo Lee, Younghyun Oh, Hyunhoe An, Hyebhin Yoon, Karl J. Friston, Seok Jun Hong, Choong-Wan Woo","doi":"10.1038/s42256-026-01296-8","DOIUrl":"10.1038/s42256-026-01296-8","url":null,"abstract":"Despite remarkable advances in artificial intelligence (AI), building agents that can autonomously pursue goals while adapting to continuously changing environments remains a fundamental challenge. Living organisms naturally exhibit such adaptive autonomy, offering a compelling source of inspiration for how artificial systems might achieve similar capabilities. Here we focus on interoception—the process of monitoring and regulating internal bodily states to maintain internal homeostasis—which underwrites an organism’s survival. Developing interoceptive AI requires abstracting internal states from their biological instantiation into functional and mathematical representations that are applicable to artificial systems. This requires explicit factorization of state variables representing internal and external environments, together with mathematical formalization of life-inspired properties governing internal-state dynamics. Importantly, internal states can also function as universally available and intrinsically valuable contexts, serving as stable reference signals that modulate learning and behaviour under changing external environments. In addition, interoception-related biological processes, such as neuromodulatory mechanisms, can be incorporated to support context-dependent adaptive behaviour. Overall, this Perspective presents a unified account on how interoception can enhance autonomy and adaptivity in artificial agents by integrating insights from cybernetics with recent advances in theories of life, reinforcement learning and neuroscience. Lee, Oh et al. propose interoception as a biologically inspired framework for building more autonomous and adaptive AI agents, learning from living organisms to build autonomous and adaptive intelligence.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 9","pages":"1335-1346"},"PeriodicalIF":29.8,"publicationDate":"2026-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148987150","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A knowledge-driven framework for predicting single-cell responses for unprofiled drugs","authors":"Jinghao Feng, Ziheng Zhao, Xiaoman Zhang, Mingfei Liu, Jingyi Chen, Xingran Quan, Boyang Fu, Jian Zhang, Yanfeng Wang, Ya Zhang, Weidi Xie","doi":"10.1038/s42256-026-01286-w","DOIUrl":"10.1038/s42256-026-01286-w","url":null,"abstract":"Predicting cellular response to chemical perturbations is critical to build virtual cells, yet experimentally profiled compounds cover only a small fraction of this space. Existing models struggle to generalize to unprofiled compounds, as they typically treat drugs as isolated identifiers without encoding their mechanistic relationships. Here, we present MAP, a framework that integrates structured biological knowledge into cellular perturbation modelling and supports zero-shot prediction for small molecules with scarce or absent profiles. (1) We construct MAP-KG, a knowledge graph that unifies 14 public resources, spanning 187,089 drugs, 22,924 genes and 694,246 mechanistic relationships. (2) We propose a knowledge-driven pretraining strategy that aligns molecular structures, protein sequences and textual mechanistic descriptions into a unified embedding space, producing mechanism-aware and transferable gene and compound embeddings. These representations are then coupled with a pretrained single-cell foundation model to condition perturbation response prediction. (3) We evaluate MAP under two zero-shot generalization regimes: unseen cell type–drug combinations and a stricter setting of unprofiled drugs, where it improves the top-50 differentially expressed gene Pearson delta correlation by up to +12.3% and +11.8%, respectively, over the strongest baselines across three benchmarks. We further perform pathway-level functional analysis via gene set enrichment analysis for in silico screening, where MAP predicts mechanism-consistent programmes on unprofiled candidate drugs, and prioritizes four out of five approved anti-cancer drugs in A-549 (non-small-cell lung cancer). Feng et al. introduce MAP, an artificial intelligence framework that integrates biological mechanism knowledge to predict how cells respond to chemical perturbation, improving generalization to untested drugs and prioritizing cancer drug candidates in virtual screening.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 9","pages":"1478-1491"},"PeriodicalIF":29.8,"publicationDate":"2026-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-026-01286-w.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148853492","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"The epistemic debt of generative AI","authors":"Irene Unceta, Paula Subías-Beltrán, Oriol Pujol","doi":"10.1038/s42256-026-01294-w","DOIUrl":"10.1038/s42256-026-01294-w","url":null,"abstract":"When authors use generative AI in cognitive tasks, without spending time and effort to understand the output, a gap opens between what they present and what they can defend. This gap widens as further work is built on top, resulting in epistemic debt.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 9","pages":"1328-1330"},"PeriodicalIF":29.8,"publicationDate":"2026-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148987151","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Multi-resolution enhancement for full-spectrum neural representations","authors":"Yuan Ni, Zhantao Chen, Shizhou Xu, Cheng Peng, Rajan Plumley, Chun Hong Yoon, Jana B. Thayer, Joshua J. Turner","doi":"10.1038/s42256-026-01287-9","DOIUrl":"10.1038/s42256-026-01287-9","url":null,"abstract":"Scientific data acquisition continues to outpace storage and analysis capabilities, making voxel-based representations increasingly intractable. Implicit neural representations (INRs) offer a promising solution by encoding signals through coordinate-based neural networks, serving as surrogates of data, with computational and storage requirements scaling with network complexity rather than data dimensionality. However, smaller INRs struggle to faithfully represent multiscale structures, high-frequency information and fine textures that constitute a large proportion of scientific measurements. We propose WIEN-INR, a theoretically guided hierarchical INR framework that distributes modelling across resolution scales and enables improved representation capacity through a novel enhancement network to recover subtle details. This multiscale architecture allows smaller networks to retain the full spatial-frequency content of the signal as well as preserve training efficiency and lower storage cost. Evaluated on distinct raw experimental measurements across scales and complexities, WIEN-INR represents a practical step towards a broader adoption of neural representations in scientific workflows, delivering compact, robust and high-fidelity representations. Ni et al. present WIEN-INR, an implicit neural representation for scientific data compression. It operates in the multiscale wavelet domain to improve compression as well as preserve fine details and signal fidelity.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 9","pages":"1492-1505"},"PeriodicalIF":29.8,"publicationDate":"2026-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-026-01287-9.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148853490","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Harnessing implicit neural representations for scientific data compression","authors":"Shuhang Gu, Kexuan Shi","doi":"10.1038/s42256-026-01290-0","DOIUrl":"10.1038/s42256-026-01290-0","url":null,"abstract":"Implicit neural representations can turn massive scientific measurements into compact functions. A hierarchical approach now helps these representations preserve richer fine-scale detail.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 9","pages":"1333-1334"},"PeriodicalIF":29.8,"publicationDate":"2026-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148987152","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A roadmap for end-to-end task-agnostic exoskeleton control","authors":"Max K. Shepherd, Ethan B. Schonhaut, Keaton L. Scherpereel, Fatima Mumtaza Tourk, Aaron J. Young","doi":"10.1038/s42256-026-01297-7","DOIUrl":"10.1038/s42256-026-01297-7","url":null,"abstract":"Lower-limb exoskeletons hold tremendous potential for augmenting human mobility across diverse applications, from assisting individuals with ageing-related mobility decline and disabilities to preventing workplace injuries in industrial settings. While human-in-the-loop optimization has driven many successful task-specific demonstrations, traditional control paradigms relying on discrete task classification face fundamental limitations: human movement is inherently continuous and infinitely variable, making task-specific approaches intractable for real-world deployment. This Perspective examines the shift towards end-to-end, artificial-intelligence-driven control systems that leverage real-time estimates of physiological state—specifically, biological joint moments. We discuss both recent advances and the key remaining challenges, including the need to optimize these systems, incorporate safety mechanisms and reduce the burden of collecting a suitable training dataset. Addressing these challenges while maintaining the flexibility and robustness of data-driven approaches will determine whether next-generation exoskeleton systems can meaningfully augment human mobility across diverse populations, environments and devices. Shepherd et al. propose end-to-end AI control of lower-limb exoskeletons based on real-time estimates of physiological signals.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 9","pages":"1347-1357"},"PeriodicalIF":29.8,"publicationDate":"2026-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148987153","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Quantitative and interface-aware prediction of peptide–protein interactions by VITAL","authors":"Wei-Hao Chen \u0000 (, ), Qi-Wen Wang \u0000 (, ), Zhi-Yi Li \u0000 (, ), Song-Yang Li \u0000 (, ), Chen Lin \u0000 (, ), Zhi-Liang Ji \u0000 (, )","doi":"10.1038/s42256-026-01291-z","DOIUrl":"10.1038/s42256-026-01291-z","url":null,"abstract":"The quantitative characterization of peptide–protein interactions remains a persistent challenge in computational biology, limiting mechanistic insight and therapeutic development. Here we introduce VITAL (Versatile InTerface-Aware global-local co-Learning), a deep learning framework built on a dual-channel co-learning architecture that fuses protein language model embeddings with a geometry-aware encoder capturing spatial proximity at the interface. This integrated learning strategy enables the simultaneous prediction of peptide–protein interactions, interface mapping and quantitative affinity estimation. We report that VITAL advances the field through three key contributions. First, VITAL reaches state-of-the-art accuracy (max AUC = 0.87) across diverse benchmarks, with experimental validation of selected novel interactions in a real-world peptide discovery attempt. Second, we demonstrate near-residue-resolution interface mapping (>60% precision), which facilitates systematic classification of four fundamental binding modes. Finally, we present the binding strength unit, a continuous affinity metric that correlates significantly with experimental data. By unifying high-resolution structural profiling with quantitative scoring, VITAL establishes a scalable computational platform that accelerates the transition from qualitative screening to quantitative, structure-informed analysis. This streamlines the discovery and design of peptide-based therapeutics. The model is publicly available as an open-source package and interactive web server to promote community adoption and further methodological innovation. Chen, Wang, Li et al. introduce VITAL, a dual-channel deep learning framework that co-learns sequence and structural contexts to quantitatively predict peptide–protein interactions, map binding interfaces and estimate affinity.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 9","pages":"1381-1395"},"PeriodicalIF":29.8,"publicationDate":"2026-08-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148769341","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}