This page is brought to you by Brian Golding (Golding@McMaster.CA) and is copied locally here to speed your access. To go to the original page (should you find something interesting or should you wish to follow links) click on

Current Issue of Nature


Volume 645 Issue 8081, 18 September 2025
Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

Volume 645 Issue 8081, 18 September 2025

Self-help

Large language models (LLMs) tend to be better at solving problems if they can be trained to set out the steps they take as they try to reach the solution. This kind of ‘reasoning’ is similar to how humans approach more complex problems, but it presents significant challenges for artificial intelligence, requiring human intervention to add labels and annotations. In this week’s issue, researchers at DeepSeek reveal how they trained a model not only to reason in this way but also to do so with minimal human input. The model DeepSeek-R1 was trained using reinforcement learning in which the model was rewarded with a high score when it solved mathematical problems correctly and penalized when it got the answer wrong. As a result, it learnt that reasoning — tackling the problem in a stepwise manner and revealing those steps — was more likely to lead to the correct answer. This led DeepSeek-R1 to self-verify and self-reflect, checking its performance before giving answers to new questions and thereby improving its performance in coding and graduate-level science problems.

Cover image: Nik Spencer/ Nature.

This Week

Top of page ⤴

News in Focus

Top of page ⤴

Books & Arts

Top of page ⤴

Opinion

Top of page ⤴

Work

Top of page ⤴

Research

  • News & Views

  • Perspective

    • Next-generation batteries will present different risks to conventional lithium-ion cells, emphasizing the need for efforts towards characterizing the abuse tolerance and hazards associated with next-generation battery materials over their life cycle and providing the data in the context of supporting affected professionals.

      • Chuanbo Yang
      • Avtar Singh
      • Donal P. Finegan
      Perspective
  • Articles

    • Magic state distillation is achieved with logical qubits on a neutral-atom quantum computer using a dynamically reconfigurable architecture for parallel quantum operations.

      • Pedro Sales Rodriguez
      • John M. Robinson
      • Sergio H. Cantú
      Article
    • Stable and robust topological edge modes are observed at finite temperatures in an array of 100 programmable superconducting qubits because of emergent symmetries present in the prethermal regime of this system.

      • Feitong Jin
      • Si Jiang
      • Dong-Ling Deng
      Article Open Access
    • A new artificial intelligence model, DeepSeek-R1, is introduced, demonstrating that the reasoning abilities of large language models can be incentivized through pure reinforcement learning, removing the need for human-annotated demonstrations.

      • Daya Guo
      • Dejian Yang
      • Zhen Zhang
      Article Open Access
    • DNA molecules can be programmed to autonomously carry out supervised learning in vitro, with the system learning to perform pattern classification from molecular examples of inputs and desired responses.

      • Kevin M. Cherry
      • Lulu Qian
      Article Open Access
    • A 2.5-billion-year record of oxygen isotopes in sedimentary sulfate reveals the transitional oxygenation of the Earth’s surface and provides constraints on the dynamic, lengthy co-oxygenation of Earth’s atmosphere and oceans.

      • Haiyang Wang
      • Chao Li
      • Huiming Bao
      Article
    • A modelling study based on satellite observations, machine learning and a chemical transport model quantifies the global and regional exposure to particulate-matter pollution and the human health impacts related to the 2023 Canadian wildfires.

      • Qiang Zhang
      • Yuexuanzi Wang
      • Kebin He
      Article Open Access
    • The discovery of chemosynthesis-based benthic communities at depths of 5,800 m to 9,533 m in the Kuril–Kamchatka and western Aleutian trenches challenges traditional perspectives on the energy sources sustaining hadal fauna.

      • Xiaotong Peng
      • Mengran Du
      • Andrey V. Adrianov
      Article Open Access
    • A study reports whole-genome sequences for 490,640 participants from the UK Biobank and combines these data with phenotypic data to provide new insights into the relationship between human variation and sequence variation.

      • Keren Carss
      • Bjarni V. Halldorsson
      • Ole Schulz-Trieglaff
      Article Open Access
    • Lithium has an essential role in the brain and is deficient early in Alzheimer’s disease, which can be recapitulated in mice and treated with a novel lithium salt that restores the physiological level.

      • Liviu Aron
      • Zhen Kai Ngian
      • Bruce A. Yankner
      Article Open Access
    • Research on Drosophila neurons shows links between the need to sleep and aerobic metabolism, indicating that the pressure to sleep may have a mitochondrial origin.

      • Raffaele Sarnataro
      • Cecilia D. Velasco
      • Gero Miesenböck
      Article Open Access
    • A study reveals a gut–brain sensory pathway through which the microbial component flagellin activates neuropod cells in the colon to signal the brain and reduce feeding in mice.

      • Winston W. Liu
      • Naama Reicher
      • Diego V. Bohórquez
      Article Open Access
    • A preclinical covalent compound, CMX410, contains a aryl fluorosulfate warhead that targets the acyltransferase domain of Mtb Pks13, an essential enzyme in cell-wall biosynthesis, making it a promising candidate for tuberculosis treatment regimens.

      • Inna V. Krieger
      • Paridhi Sukheja
      • Case W. McNamara
      Article
    • Excised signal circles are circular DNA by-products of V(D)J recombination that form a complex with the V(D)J recombinase, and when increased in abundance, result in increased mutagenesis, causing adverse outcomes in cancer.

      • Zeqian Gao
      • James N. F. Scott
      • Joan Boyes
      Article Open Access
    • N-glycans on glycoRNAs prevent innate immune sensing of endogenous small RNAs, and the natural mechanism they use demonstrates how glycoRNAs exist on the cell surface and in the endosomal network without inducing autoinflammatory responses.

      • Vincent R. Graziano
      • Jennifer Porat
      • Vijay A. Rathinam
      Article
    • A synthetic biology system called SMART has been developed that uses conditional protein splicing for the programmable ligation of functional proteins from previously defined molecular combinations on cell surfaces.

      • Christian Kofoed
      • Girum Erkalo
      • Tom W. Muir
      Article Open Access
Top of page ⤴

Amendments & Corrections

Top of page ⤴

Innovations In

  • The outlook for preventing, diagnosing and treating Alzheimer’s disease is brighter than it’s ever been.

    Innovations In
Top of page ⤴
Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing

Search

Quick links