内容来源：https://research.google/blog/google-research-2025-bolder-breakthroughs-bigger-impact/

内容总结：

【谷歌发布2025年度研究总结：AI驱动多领域突破，加速科学与社会创新】

谷歌研究院副总裁尤西·马蒂亚斯于2025年12月18日代表团队发布年度研究成果综述。过去一年中，谷歌通过“研究魔法循环”的加速运转，在人工智能、量子计算、地球科学、医疗健康等领域取得系列突破，并将前沿技术转化为实际应用。

AI技术实现跨越式发展
谷歌通过提升生成式模型的效率、事实准确性与多语言文化适应性，推出新一代大语言模型Gemini 3。该模型在多项事实性基准测试中达到领先水平，已应用于Gemini应用、AI搜索概览及Vertex AI平台。同时，团队研发的生成式UI技术可为用户动态创建交互式视觉界面，增强搜索与应用的沉浸式体验。

量子计算迈向实用化
基于超导量子比特的长期研究，谷歌团队开发的“量子回声”算法在Willow量子芯片上的运行速度比经典超级计算机快1.3万倍，为分子原子相互作用研究提供新路径，推动药物研发与核聚变能源等领域的应用探索。

科学发现进程显著提速
谷歌推出“AI协研员”多智能体系统，可帮助科学家生成新假设并编写专业级实验软件。该工具已在斯坦福大学和伦敦帝国理工学院投入应用，成功加速肝纤维化药物筛选及抗菌耐药性研究。在基因组学领域，开源工具DeepSomatic助力癌症细胞变异分析，单细胞分析模型C2S-Scale则提出关于癌细胞行为的新假设。

地球科学助力灾害应对
谷歌地球AI平台整合遥感、气象、人口动态等多维度数据，通过Gemini的推理能力生成行星级洞察。FireSat卫星星座利用AI实现近实时野火监测，可探测教室规模的火灾。洪水预测模型已覆盖150个国家超20亿人口， cyclone预测模型能将气旋路径预测提前至15天。

医疗与教育领域深度赋能
conversational医疗助手AMIE在模拟诊疗中表现优于初级保健医生，开源模型MedGemma支持医疗报告生成与电子病历解读。教育领域通过LearnLM模型实现个性化学习，试点研究表明学生知识保留率提升11个百分点。

隐私保护与算法创新并进
谷歌开源差分隐私库Jax Privacy 1.0，并发布首个完全基于差分隐私训练的开源大模型VaultGemma。在算法层面，Speech-to-Retrieval引擎实现语音直接检索，TimesFM时序预测模型月查询量达数亿次。

全球研究网络持续扩展
谷歌研究院计划于2026年在新加坡设立新研究中心，持续通过学术合作、开源发布及博士生培养项目推动全球科研生态发展。

报告强调，谷歌正通过安全可靠的技术路径，推动人工智能成为人类创造力的放大器，致力于解决全球性挑战并赋能各行各业。

中文翻译：

谷歌研究 2025：更大胆的突破，更深远的影响

2025年12月18日
谷歌研究副总裁兼负责人 Yossi Matias，代表谷歌研究团队

2025年，研究的“魔法循环”加速运转。谷歌研究团队取得了开创性突破，并将我们的研究转化为现实，对产品、科学和社会产生了深远影响。

多年来，谷歌研究团队持续投资于推进多个战略领域的研究与技术。我们的工作跨越不同时间维度：从大胆的“登月计划”和探索可能性的好奇心驱动型变革性研究，到加速产生影响的创新与应用研究。研究的“魔法循环”正在加速——我们正推动研究突破，并将其转化为现实世界的解决方案，通过与谷歌内部众多团队及全球合作伙伴的紧密协作，对产品、科学和社会产生影响。

这是成果丰硕的一年！我们的基础性人工智能突破助力生成式模型变得更高效、更准确、支持更多语言和文化，并推出了生成式用户界面。我们推进了新的架构和算法研究，开创了有助于加速科学发现的人工智能工具和智能体模型。我们在量子计算领域取得突破，使其更接近实际应用；推进了地球科学研究，实现了前所未有的行星级理解水平；推动了包括基因组学、生物学和神经科学在内的科学领域发展；并在气候韧性、健康和教育等社会优先事项上取得了进展。

推进生成式模型：更高效、更准确、支持多语言与文化

为了助力这个快速创新的时代，我们正大力投资于效率提升，使谷歌产品更具成本效益和能源效率，并为行业树立标杆。我们继续开发基于推测解码的新方法，例如块验证，以进一步加速效率提升。在基础设施栈的另一端，LAVA是一种新的调度算法，能持续重新预测虚拟机上的任务生命周期。它旨在优化大型云数据中心的资源效率，同时不牺牲可靠性。

同样关键的是，我们自2021年起在大型语言模型事实性方面的开创性研究，帮助Gemini 3成为我们迄今为止能力最强、事实性最高的模型。它在SimpleQA Verified等公共事实性基准测试，以及我们与Google DeepMind和Kaggle共同发布的新FACTS基准套件上，都达到了最先进的性能水平。用户可以确信，诸如Gemini应用、搜索中的AI概览和AI模式、Vertex AI等产品，其输出都基于真实世界知识。今年，我们研究了大型语言模型如何表达不确定性；提出了一个评估模型参数中编码的事实知识是否多于其输出所表达的框架；发布了一个名为ECLeKTic的评估跨语言知识的多语言数据集；等等。

我们还探索了检索增强生成系统中充分上下文的作用，这类系统通过提供相关的外部上下文来增强大型语言模型。我们证明了，有可能知道大型语言模型何时拥有足够的信息来正确回答问题。这项工作支持了Vertex AI RAG引擎中LLM重排序器的发布，从而带来了更好的检索指标和系统准确性。

随着多模态内容的兴起，我们将事实性研究工作扩展到了图像、音频、视频、3D环境和大型语言模型生成的应用。这项工作有助于提升谷歌视频和图像模型家族（包括Veo、Imagen和Nano Banana）的质量。这是研究循环的一个绝佳例证，展示了我们如何持续适应真实用户需求。我们最新的研究包括提高文本到图像生成和图像描述的准确性，以及创建用于评估智能体在3D环境中进行长期记忆推理能力的3DMem-Bench。

我们长期进行的多语言研究帮助Gemma扩展至超过140种语言，使其成为当今最佳的多语言开源模型。我们还在为模型增强社会文化智能，使其适应多样化的用户需求和全球背景。我们引入了TUNA，一个用户需求与行为的综合分类法；推出了一个基于社区的数据收集平台，以覆盖代表性不足的语言和地区；并开发了新方法，将模型建立在多样化的文化知识和数据集之上。这项研究有助于确保谷歌模型能够以负责任且具备文化意识的方式与全球用户建立联系。

引入生成式用户界面的交互体验

在一个用户期待更具吸引力和视觉体验的世界里，我们在Gemini 3中引入了一种新颖的生成式用户界面实现。这一强大功能使人工智能模型能够根据提示，动态创建沉浸式的视觉体验和交互界面，例如网页、游戏、工具和应用程序。我们的研究在谷歌搜索的AI模式以及Gemini应用中的动态视图等实验中得以实现。

量子计算：下一个前沿领域

我们对量子计算的战略投资，有望加速计算和科学发现的下一个前沿。20世纪80年代，Clarke、Devoret和Martinis为超导量子比特奠定了基础，这使他们荣获2025年诺贝尔物理学奖。此后的40年历程催生了新兴的量子计算产业，并带来了我们最近宣布的、发表在《自然》杂志封面上的可验证量子优势等突破。这项工作描述了我们名为“量子回声”的算法，它在我们的Willow芯片上运行，比世界上最快超级计算机上最好的经典算法快13,000倍。它提供了一种新方法来解释使用核磁共振光谱观察到的分子中原子间的相互作用。这使我们更接近量子计算的实际应用，例如推进药物设计，并助力实现聚变能源。

加速科学发现

人工智能驱动的模型和平台正在从根本上改变科学研究的进行方式。我们发布了AI联合科学家，这是谷歌研究、云AI和Google DeepMind之间的合作成果。这个多智能体人工智能系统帮助科学家生成新颖的假设。我们还分享了由人工智能驱动的经验软件系统，这是一个由Gemini支持的编码助手，帮助科学家编写专家级的经验软件来评估和迭代假设。这些工具加速了科学发现的过程本身。它们开启了一个未来：实验室里的每位科学家都将拥有一组人工智能助手，同时研究激发其研究的科学挑战的数千种潜在解决方案。在斯坦福大学，我们的AI联合科学家已帮助识别出可能被重新用于治疗肝纤维化的药物。在伦敦帝国理工学院，研究抗菌素耐药性的研究人员发现，它在几天内就产生了他们团队花费数年才得出的相同假设。

推进科学——从生物学到基因组学再到神经科学

我们持续推动核心科学研究。DeepSomatic和C2S-Scale加入了人工智能抗击癌症的行列，并为全新的疗法铺平道路。发表在《自然·生物技术》上的DeepSomatic是一个开源工具，它建立在谷歌十年基因组学研究的基础上，帮助科学家和医生识别癌细胞中的基因变异。我们在Children's Mercy的合作伙伴正在使用它来了解某种特定癌症如何以及为何影响患者，以开发个性化疗法。C2S-Scale是我们与Google DeepMind和耶鲁大学合作发布的，是一个拥有270亿参数的单细胞分析基础模型，因生成关于癌细胞行为的新颖假设而登上新闻头条。

在神经科学方面，我们在《自然》杂志上发表了首个利用普通光学显微镜全面绘制脑组织块中所有神经元及其连接的方法。我们与奥地利科学技术研究所合作，应用了我们为连接组学开发的一套图像分析和机器学习工具，利用我们十多年来在该科学领域的贡献来理解大脑的工作原理。我们希望这种名为LICONN的方法能使全球更多实验室开展连接组学研究。

我们还与HHMI Janelia和哈佛大学合作，开源了斑马鱼活动预测基准。它拥有幼年斑马鱼大脑中超过70,000个神经元的记录，将使科学家首次能够研究整个脊椎动物大脑的结构连接与动态神经活动之间的关系。

此外，我们展示了大型语言模型如何帮助我们理解人类大脑。在与普林斯顿大学、纽约大学和希伯来大学历时五年进行的一系列研究中，我们探索了人类大脑和深度语言模型处理自然语言方式的联系。我们发现，人类大脑语言和言语区域的神经活动与基于Transformer的语音转文本模型的语言和言语嵌入之间存在显著的一致性，并展示了大脑中语言处理的时间结构如何对应于深度语言模型的分层结构。我们的研究表明，深度学习模型中的语言表征可能为理解大脑的神经编码提供一个新颖的框架；同时也为创建具有更好信息处理能力的人工神经网络开辟了创新途径。

赋能行星智能与危机韧性

谷歌地球AI计划是一项雄心勃勃的研究，汇聚了多方努力以产生更大影响。它由谷歌内部多个团队合作开发，建立在我们多年对世界进行建模的基础上，结合Gemini的先进推理能力，提供了对我们星球前所未有的理解水平。它汇集了谷歌的许多地理空间模型和技术，如遥感图像、天气、空气质量、洪水、人口动态、AlphaEarth Foundations、移动性、地图等。得益于Gemini的推理能力，地球AI能够在几分钟内合成关于地球的海量数据集，生成以往需要数年研究才能获得的洞见。地球AI产品已在谷歌地图平台、谷歌地球中提供，并通过谷歌云提供给可信测试者，并已被合作伙伴用于从城市规划到灾难响应的关键任务。

我们在为理解地球提供AI能力的气候模型方面也取得了重大进展，帮助社区准备和应对恶劣天气与自然灾害。今年，我们与Earth Fire Alliance、Moore基金会和Muon Space合作，发射了FireSat星座的第一颗卫星。FireSat被《时代》杂志评为2025年最佳发明之一，它利用人工智能为应急响应人员提供关键的近实时洞察。它已经探测到其他天基系统未发现的小型野火，当完全部署超过50颗卫星后，将能够探测到地球上任何地方教室大小的野火。

我们还扩展了洪水预测模型，覆盖150个国家超过20亿人口，应对最重大的河流洪水事件，帮助社区保持安全并获取信息。我们与Google DeepMind的同事合作，推出了一个使用随机神经网络进行气旋预测的实验模型，帮助气象机构提前多达15天预测气旋路径。此外，我们与Google DeepMind合作推出了WeatherNext 2，它提供了我们迄今为止最准确的中期人工智能天气预报。现在，搜索、Gemini和Pixel Weather的用户，以及谷歌地图和谷歌云上的开发者都可以使用它。

今年年初，我们将搜索中的临近预报扩展到非洲，首次为整个非洲大陆的用户提供高度精确的短期降水预报。此后，我们已将其提供给全球用户。它由我们的MetNet模型驱动，是搜索中首个在全球范围内运行的人工智能天气模型。在印度，芝加哥大学和印度农业与农民福利部使用谷歌的NeuralGCM模型，向3800万农民发送了更长范围的季风预报，帮助他们做出关于种植什么和何时种植的关键决策。

推进健康人工智能

随着我们取得可能显著改革医疗保健的科学突破，我们正与合作伙伴和医疗专业人士合作，负责任地将新能力带给全球人民。AMIE是我们与Google DeepMind共同开发并在《自然》杂志上发表的对话式医疗智能体。它现在能够通过多模态证据进行推理，支持纵向疾病管理，在模拟环境中与专业患者演员互动时，其表现达到或优于初级保健医生。我们正在探索这项研究如何能够实现一种以医生为中心、对AMIE进行异步监督的模式。我们还向部分选择加入的用户推出了Plan for Care Lab，这是Fitbit最新的实验性功能。它旨在帮助用户在家评估症状和为即将到来的医生就诊做准备时，获得个性化支持。此外，MedGemma——谷歌最强大的多模态医学理解开源模型，已作为我们健康人工智能开发者基础的一部分提供。它可以支持分类、报告生成或解读复杂的电子健康记录等任务，对医学研究和产品开发非常有用。自发布以来，MedGemma和健康人工智能开发者基础已获得超过200万次下载。另外，我们的Open Health Stack在世界经济论坛上获得认可，因其有助于解决医疗可及性的不平等问题。它为开发人员提供了构建模块，以创建用于资源匮乏环境下的下一代数据驱动型医疗应用。

推进学习与教育

Gemini现已融入LearnLM，这是谷歌去年宣布的为学习而微调的模型家族。我们在Google Labs上推出了“按你的方式学习”，由LearnLM的基础能力驱动。它通过生成源材料的多种引人入胜的呈现形式，探索教科书的未来。它将静态教科书转变为为每个学生量身定制的主动学习体验，配有互动测验，可实现实时评估、反馈和内容个性化。在我们的有效性研究中，使用它的学生在记忆测试中的得分高出11个百分点。我们还在加纳的数千名高中生中试点使用我们的LearnLM模型进行答案评估。此外，我们通过以学习者为中心的方法探索了教育与健康的交叉点，量化了LearnLM在医学教育环境中的益处。

这项研究使我们更接近实现一个未来：人工智能让每个人的学习都更有效。我们与谷歌内部多个团队合作，发表了《人工智能与学习的未来》，分享了我们在学习科学基础上负责任地赋能人工智能学习的方针。我们正在创造个性化的教学体验，赋能教育工作者，并努力应对批判性思维和平等获取等挑战。

与此同时，我们的人工智能素养工作旨在激励下一代创新者。与斯坦福学习加速器合作推出的AI探索，让学生们能够扮演谷歌研究员的角色，使用人工智能解决洪水预测和检测眼疾等挑战。在计算机科学教育周期间，数百名谷歌志愿者将这些探索带到了世界各地的课堂。

推进机器学习基础与算法研究

我们广泛的机器学习基础和算法研究是跨领域突破性进展的基石。这项工作为产品和服务提供了核心框架，并支撑着下一代模型和智能系统的开发。例如，我们通过新的语音到检索引擎改进了语音搜索，该引擎可直接解读和检索语音查询中的信息，而无需先将其转换为文本。我们对丰富人类反馈的最先进预测建模，提升了产品中的文本到图像生成质量，包括Imagen3、谷歌广告中的创意生成和编辑，以及购物虚拟试穿。我们还将这项研究扩展到提升拉斯维加斯Sphere影院《绿野仙踪》电影发布中的视频生成质量。

我们算法研究的影响远远超出了谷歌产品。我们的TimesFM模型帮助企业进行时间序列预测，现在在BigQuery和AlloyDB中每月处理数亿次查询。我们引入了一种使用上下文微调的新方法，它教会模型如何在推理时从多个示例中学习，以进一步提升其性能。我们的移动性人工智能模型利用了我们在地图和交通领域二十年的创新，为交通机构提供强大的工具，用于数据驱动的政策制定和交通管理。它可以理解交通和停车模式，模拟系统以允许工程师测试不同场景，并为交通网络识别有效的解决方案。这补充了我们在谷歌地图和搜索中面向消费者的突破，例如用于计算预计到达时间和优化行程规划的专用模型。

此外，我们探索了经济学和计算领域的诸多主题，从模块化市场中的定价动态、采购拍卖，到数据驱动的机制设计以及优化广告拍卖的各种方法。我们还研究了博弈中的交换遗憾和相关均衡。

随着人工智能日益融入日常生活，将其构建在以隐私为核心的基础上对用户和行业至关重要。为此，我们开发并发布了用于隐私学习和隐私分析的新算法，并开源了强大的软件工具以实现外部可验证性。例如，我们推出了Parfait，这是一个面向企业和开源项目的新GitHub组织。它支持了谷歌从Gboard到谷歌地图的联邦学习和分析部署。我们还宣布了Jax Privacy 1.0，这是一个用于差分隐私机器学习的库，我们用它训练了VaultGemma——这是迄今为止规模最大、能力最强、从头开始使用差分隐私训练的开源模型，其权重已在Hugging Face和Kaggle上提供。通过提升我们的隐私能力，我们为企业和用户提供了更强的保护。

引入新颖架构

我们的基础机器学习研究引入了先进方法以开启新机遇。嵌套学习是一种新的机器学习范式，它代表了我们对深度学习理解的一次飞跃。它将模型架构和优化视为一个单一系统，其中包含几个更小的嵌套优化问题。通过统一这些元素，它解决了灾难性遗忘的问题，即大型语言模型在学习新任务后变得健忘，对旧任务能力下降。这项研究可能帮助我们构建下一代能力更强、能够自我改进的人工智能。同时，我们的Titans架构和MIRAS框架标志着序列建模的重大进步。它们通过采用在数据输入时学习记忆的深度神经网络，使人工智能模型工作速度更快，并能处理海量上下文，从而改善了人工智能的长期记忆。

我们还引入了MUVERA，这是一种新颖的检索算法，它将复杂的多向量检索简化为单向量最大内积搜索，以显著提升的效率实现了最先进的性能。它为信息检索创造了新的可能性，可用于推荐系统和自然语言处理等应用。我们在图基础模型方面的进展则推动了图学习的前沿。虽然大多数图神经网络都固定在模型训练所基于的特定图上，但我们开发了能够泛化到任意表格、特征和任务的图基础模型。这为模型复用开辟了新途径。

与研究生态系统合作

我们与全球学术界、行业领导者、政府和科研机构合作。我们也继续通过从山景城到东京、悉尼和波兰的Research@活动与生态系统互动，并通过谷歌奖学金项目支持数百名博士生。

作为一个全球团队，我们继续将足迹扩展到主要中心之外。在巩固了我们在非洲（阿克拉和内罗毕）的研究投资和创新以及在澳大利亚的存在后，我们现在正准备于2026年在新加坡开设一个新的谷歌研究中心。

我们通过出版物、会议、学术讲座、基准测试、数据集和开源发布来分享我们的工作。我们在会议上赞助和主办研讨会，最近一次是在NeurIPS。我们最近引入了一个实验性项目，在科学家提交会议论文进行同行评审之前，为他们提供自动化反馈，帮助他们严格验证工作并加速研究流程。此外，我们与NotebookLM合作推出了谷歌研究精选笔记本，以使更广泛的社区更容易获取研究成果。

人工智能：人类智慧的放大器

这是研究的黄金时代。技术突破和科学进步从未如此迅速地转化为有影响力的现实解决方案，而这些解决方案反过来又带来了新的数据和问题，激发了基础研究的新途径。这个“魔法循环”正在显著加速，其驱动力是更强大的模型、支持科学发现的新智能体工具，以及开放的平台和工具。

我们与谷歌的同事和合作伙伴一起，正在推进旨在为不同领域提供帮助的研究和技术。我们的研究立足于对安全与信任的严谨承诺，旨在释放人类潜能——无论是帮助科学家加速研究，还是帮助学生更有效地学习和掌握新概念，抑或是赋能医生、开发者或教师。

这确实是从事研究的激动人心的时刻。我们能够利用谷歌人工智能基础设施、模型、平台和世界级人才的全栈优势，并为数十亿人使用的产品做出贡献。我们将继续在我们的传统基础上再接再厉，提出当今最大的问题，并致力于促成未来的解决方案。我们将继续以大胆且负责任的方式推进人工智能，造福社会，帮助提升人类能力，使人工智能成为人类智慧的放大器。

致谢

感谢谷歌研究的每一位成员，以及许多协作者，他们为这篇博客及其所代表的工作做出了贡献。

英文来源：

Google Research 2025: Bolder breakthroughs, bigger impact
December 18, 2025
Yossi Matias, Vice President & Head of Google Research, on behalf of the Google Research team
In 2025, the magic cycle of research accelerated. Google Research teams delivered pioneering breakthroughs and brought our research to reality with impact on products, science and society.
Google Research teams have invested over the years in advancing research and technology in a diverse range of strategic areas. We are working across time horizons, from bold moonshots and curiosity-driven transformative research where we explore the art of the possible, to innovation and applied research with accelerated impact. The Magic Cycle of research is accelerating — we’re driving research breakthroughs and translating them into real-world solutions, with impact on products, science and society, in close collaboration with many teams across Google and global partners.
This was quite a year! Our foundational AI breakthroughs helped make generative models more efficient, factual, multilingual, and multi-cultural, and we introduced generative UI. We advanced new architectures and algorithmic research and pioneered AI tools and agentic models that help accelerate scientific discovery. We achieved quantum breakthroughs that bring us closer to real-world applications of quantum computing; advanced research on Earth sciences to enable a level of planetary understanding never before possible; drove forward scientific domains including genomics, biology and neuroscience; and made headway on societal priorities like climate resilience, health and education.
Advancing generative models to be more efficient, factual, multilingual and multi-cultural
To help fuel this era of rapid innovation, we’re investing in efficiency, making Google products more cost and energy efficient, and setting the bar for the industry. We continue to develop new approaches based on speculative decoding, such as block verification, to further accelerate efficiency gains. At the other end of the infrastructure stack, LAVA is a new scheduling algorithm that continuously re-predicts the lifespans of tasks on virtual machines. It is designed to optimize resource efficiency in large cloud data centers, without sacrificing reliability.
Equally critical, our pioneering research on LLM factuality, dating back to 2021, helps make Gemini 3 our most capable and factual LLM yet. It achieves state-of-the-art performance on public factuality benchmarks like SimpleQA Verified and the new FACTS benchmark suite that we released with Google DeepMind and Kaggle. Users can be confident that products such as the Gemini app, AI Overviews and AI Mode in Search, and Vertex AI all provide outputs grounded in world knowledge. This year we studied how LLMs convey uncertainty; presented a framework for assessing whether LLMs encode more factual knowledge in their parameters than they express in their outputs; presented a multilingual dataset that evaluates cross-lingual knowledge, called ECLeKTic; and more.
We also explored the role of sufficient context in retrieval augmented generation systems, which enhance LLMs by providing them with relevant external context. We demonstrated that it is possible to know when an LLM has enough information to provide a correct answer to a question. This work supported the launch of the LLM Re-Ranker in the Vertex AI RAG Engine, leading to better retrieval metrics and system accuracy.
With the rise of multimodal content, we’ve expanded our work on factuality to images, audio, video, 3D environments and LLM-generated applications. This work helps to improve the quality of Google’s video and image model families, including Veo, Imagen and Nano Banana. It is a great example of the cycle of research and how we’re continuously adapting to real user needs. Our latest research includes making text-to-image generation and image captions more accurate, and creating 3DMem-Bench for evaluating an agent’s ability to reason over long-term memory in 3D.
Our long-running multilinguality research helped Gemma expand to over 140 languages, making it today’s best multilingual open model. We’re also augmenting our models with socio-cultural intelligence, attuning them to diverse user needs and global contexts. We introduced TUNA, a comprehensive taxonomy of user needs and actions, launched a community-based data collection platform to target under-represented languages and geographies, and developed new methods to ground models in diverse cultural knowledge and datasets. This research helps to ensure that Google models can connect with users globally in responsible and culturally-aware ways.
Introducing interactive interfaces with generative UI
In a world where users expect more engaging and visual experiences, we introduced a novel implementation of generative UI in Gemini 3. This powerful capability enables AI models to dynamically create immersive visual experiences and interactive interfaces, such as web pages, games, tools and apps, in response to a prompt. Our research comes to life in AI Mode on Google Search, and in experiments such as dynamic view, in the Gemini app.
Quantum computing: The next frontier
Our strategic investment in quantum computing is poised to accelerate the next frontier of computing and scientific discovery. In the 1980s, Clarke, Devoret, and Martinis laid the foundations for superconducting qubits, which led to their recognition as 2025 Physics Nobel Laureates. The 40-year journey since has yielded the nascent quantum computing industry and led to breakthroughs like our recently announced verifiable quantum advantage, published on the cover of Nature. This work describes our “Quantum Echoes” algorithm, which runs on our Willow chip 13,000 times faster than the best classical algorithm on one of the world’s fastest supercomputers. It offers a new way to explain interactions between atoms in a molecule observed using nuclear magnetic resonance spectroscopy. It brings us closer to real-world applications of quantum computing, such as advancing drug design and helping to make fusion energy a reality.
Accelerating scientific discovery
AI-powered models and platforms are fundamentally changing how science is conducted. We released AI co-scientist, a collaboration across Google Research, Cloud AI and Google DeepMind. This multi-agent AI system helps scientists generate novel hypotheses. We also shared our AI-powered empirical software system, a Gemini-backed coding agent to help scientists write expert-level empirical software to evaluate and iterate on hypotheses. These tools accelerate the very process of making scientific discoveries. They open the door to a future where every scientist in a lab has a team of AI assistants simultaneously investigating thousands of potential solutions to the scientific challenges that motivate their research. Already at Stanford, our AI co-scientist has helped identify drugs that could be repurposed to treat liver fibrosis. At Imperial College London, researchers working on antimicrobial resistance found that it produced the same hypothesis in days that their team took years to develop.
Advancing science — from biology to genomics to neuroscience
We continue to advance core scientific research. DeepSomatic and C2S-Scale join the AI-powered fight against cancer and are paving the way for brand-new therapies. Published in Nature Biotechnology, DeepSomatic is an open-source tool that builds on 10 years of genomics research at Google and helps scientists and doctors identify genetic variants in cancer cells. Our partners at Children’s Mercy are using it to understand how and why a particular form of cancer affects a patient in order to develop personalized cures. C2S-Scale, which we released in collaboration with Google DeepMind and Yale, is a 27 billion parameter foundation model for single-cell analysis that made headlines for generating a novel hypothesis about cancer cellular behavior.
Turning to neuroscience, we published in Nature the first-ever method for using commonly available light microscopes to comprehensively map all the neurons and their connections in a block of brain tissue. Working with the Institute of Science and Technology Austria, we applied our suite of image analysis and ML tools for connectomics, leveraging over a decade of contributions we’ve made to this scientific field to understand the workings of the brain. We hope the method, called LICONN, will enable more labs around the world to pursue connectomics studies.
We also open-sourced the Zebrafish Activity Prediction Benchmark (ZAPBench) in collaboration with HHMI Janelia and Harvard. With recordings of more than 70,000 neurons from the larval zebrafish brain, it will enable scientists to investigate the relationship between the structural wiring and dynamic neural activity across an entire vertebrate brain for the first time.
Plus, we demonstrated how LLMs can help us understand the human brain. In a series of studies conducted over five years with Princeton University, NYU, and HUJI, we explored connections in the ways the human brain and deep language models process natural language. We discovered remarkable alignment between the neural activity in the speech and language areas of the human brain and the speech and language embeddings of a Transformer-based speech-to-text model, and showed how the temporal structure of language processing in the brain corresponds to the layered hierarchy of deep language models. Our research indicates that language representation in deep learning models could offer a novel framework for understanding the brain’s neural code; it also paves the way for innovative approaches to creating artificial neural networks with better information processing capabilities.
Enabling planetary intelligence and crisis resilience
The Google Earth AI initiative is ambitious research that converges multiple efforts for greater impact. Developed in collaboration with teams across Google, it builds on our years of modeling the world, paired with Gemini’s advanced reasoning, to offer an unprecedented level of understanding about our planet. It brings together many of Google’s geospatial models and technologies such as remote sensing imagery, weather, air quality, floods, population dynamics, AlphaEarth Foundations, mobility, maps and more. Thanks to Gemini’s reasoning power, Earth AI can synthesize vast datasets about the planet to generate insights in minutes that would previously take years of research. Earth AI offerings are available in Google Maps Platform, Google Earth and to Trusted Testers via Google Cloud, and are already being used by partners, helping cities, enterprises and nonprofits with critical tasks from urban planning to disaster response.
We’ve also made significant strides with the climate models that feed our AI capabilities for understanding the Earth, helping communities to prepare for and respond to severe weather and natural disasters. This year, in collaboration with the Earth Fire Alliance, the Moore Foundation and Muon Space, we launched the first satellite in the FireSat constellation. Named as one of TIME magazine’s best inventions of 2025, FireSat uses AI to provide critical near–real-time insights for first responders. It has already detected small wildfires not caught by other space-based systems, and when fully operational with over 50 satellites, it will be able to detect a classroom-sized wildfire anywhere on Earth.
We also expanded our flood forecasting models to cover over 2 billion people in 150 countries for the most significant riverine flood events, helping communities stay safe and informed. We partnered with our colleagues at Google DeepMind to debut an experimental model for cyclone predictions using stochastic neural networks that's helping weather agencies predict a cyclone’s path up to 15 days in advance. Moreover, we collaborated with Google DeepMind to launch WeatherNext 2, which delivers our most accurate, mid-range AI weather forecasts to date. It’s now available to users of Search, Gemini and Pixel Weather as well as to developers on Google Maps and Google Cloud.
At the start of the year, we expanded Nowcasting on Search to Africa, bringing highly precise, short-term precipitation forecasts to users across the continent for the first time. We have since made this available for users worldwide. Powered by our MetNet model, it represents the first AI weather model on Search to operate at a global scale. In India, the University of Chicago and the Indian Ministry of Agriculture and Farmers’ Welfare used Google’s NeuralGCM model to send longer-range monsoon forecasts to 38 million farmers, helping them make critical decisions about what to plant and when.
Advancing Health AI
As we make scientific breakthroughs with the potential to significantly reform healthcare, we’re working with partners and healthcare professionals to bring new capabilities responsibly to people around the world. AMIE is our conversational medical agent developed together with Google DeepMind and published in Nature. It can now reason through multimodal evidence and support longitudinal disease management as well as or better than primary care physicians under simulated settings with professional patient actors. We’re exploring how this research could enable a physician-centered model with asynchronous oversight of AMIE. We also launched Plan for Care Lab, Fitbit’s latest experimental capability, to a select number of opt-in users. It’s designed to help users access personalized support when assessing symptoms at home and preparing for an upcoming doctor’s visit. In addition, MedGemma, Google's most capable open model for multimodal medical comprehension, is available as part of our Health AI Developer Foundations (HAI-DEF). It can support tasks such as classification, report generation, or interpreting complex electronic health records, making it useful for medical research and product development. Since launch, MedGemma and HAI-DEF have >2M downloads. Plus, our Open Health Stack was recognized at the World Economic Forum for helping to address inequities in health access. It provides the building blocks for developers to create next-gen, data-driven healthcare apps for use in low-resource settings.
Advancing learning and education
Gemini is now infused with LearnLM, Google’s family of models fine-tuned for learning, announced last year. We launched Learn Your Way on Google Labs, powered by LearnLM’s foundational capabilities. It explores the future of textbooks by generating multiple engaging representations of the source material. It transforms static textbooks into active learning experiences that are tailored for every student, with interactive quizzes that enable real-time assessment, feedback, and content personalization. In our efficacy study, students using it scored 11 percentage points higher on retention tests. We also piloted our LearnLM model for answer assessment with thousands of high school students in Ghana. Plus, we explored the intersection of education and health through a learner-centric approach quantifying the benefits of LearnLM in medical education settings.
This research brings us closer to realizing a future where AI makes learning more effective for everyone. In collaboration with teams across Google, we published “AI and the Future of Learning”, sharing our approach, grounded in learning science, to responsibly enable AI for learning. We’re creating personalized teaching experiences, empowering educators, and working to address challenges such as critical thinking and equal access.
In parallel, our AI Literacy efforts aim to inspire the next generation of innovators. AI Quests, launched with the Stanford Accelerator for Learning, allows students to step into the shoes of Google researchers and use AI to solve challenges like flood forecasting and detecting eye disease. During Computer Science Education Week, hundreds of Googler volunteers brought these quests to classrooms around the world.
Advancing ML foundations and algorithmic research
Our broad foundational ML and algorithmic research is the bedrock for groundbreaking advances across domains. This work provides the essential frameworks that power products and services, and underpins the development of next-generation models and intelligent systems. We improved voice search, for example, with our new Speech-to-Retrieval engine, which directly interprets and retrieves information from a spoken query without having to convert it first to text. And our state-of-the-art predictive modeling of rich human feedback improved text-to-image generation quality in products, including Imagen3, creative generation and editing in Google Ads, and virtual try on for shopping. We also extended this research to improve video generation quality in the Wizard of Oz film launch at Sphere in Las Vegas.
The impact of our algorithmic research extends well beyond Google products. Our TimesFM model, which helps businesses with time-series forecasting, now has hundreds of millions of queries per month in BigQuery and AlloyDB. We introduced a novel approach using in-context fine-tuning, which teaches the model how to learn from multiple examples at inference time to further enhance its performance. Our Mobility AI model leverages our two decades of innovation in maps and transportation to provide transportation agencies with powerful tools for data-driven policymaking and traffic management. It can understand traffic and parking patterns, simulate systems to allow engineers to test different scenarios, and identify effective solutions for transportation networks. This complements our consumer-facing breakthroughs in Google Maps and Search, such as specialized models for calculating ETAs and optimizing trip planning.
Additionally, we’ve explored a range of topics in economics and computation from pricing dynamics in modular marketplaces and in procurement auctions, to data-driven mechanism design and various approaches to optimize ad auctions. We also studied swap regret and correlated equilibria in games.
As AI becomes increasingly integrated into our daily lives, building it with privacy at its core is critical for users and industries. To this end, we’ve developed and published novel algorithms for private learning and private analytics, and open sourced robust software tools to enable external verifiability. For example, we introduced Parfait, a new GitHub organization for businesses and open-source projects. It has supported Google deployments of federated learning and analytics from Gboard to Google Maps. We also announced Jax Privacy 1.0, a library for ML with differential privacy, which we used to train VaultGemma, the largest and most capable open model trained from scratch with differential privacy, with weights available on Hugging Face and Kaggle. By leveling up our privacy capabilities, we offer much stronger protections to businesses and users
Introducing novel architectures
Our foundational ML research introduces advanced approaches to enable new opportunities. Nested Learning is a new ML paradigm that represents a leap forward in our understanding of deep learning. It treats model architecture and optimization as a single system that contains several, smaller, nested optimization problems. By unifying these elements, it solves the problem of catastrophic forgetting, when LLMs become forgetful and less capable at old tasks after learning new tasks. This research could help us build the next generation of more capable, self-improving AI. Meanwhile, our Titans architecture and the MIRAS framework mark a significant advancement in sequence modelling. They allow AI models to work much faster and handle massive contexts by employing deep neural networks that learn to memorize as data comes in, improving AI’s long-term memory.
We also introduced MUVERA, a novel retrieval algorithm that reduces complex multi-vector retrieval back to single-vector maximum inner product search, achieving state-of-the-art performance with significantly improved efficiency. It creates new possibilities for information retrieval for use in applications such as recommendation systems and natural language processing. And our progress on graph foundational models pushes the frontiers of graph learning. While most graph neural networks are fixed to a specific graph on which the model has been trained, we developed graph foundational models capable of generalizing to arbitrary tables, features and tasks. This opens up new avenues for model reuse.
Collaborating with the research ecosystem
We partner with the academic community, industry leaders, governments and scientific institutes around the world. We also continue to engage the ecosystem through our Research@ events from Mountain View to Tokyo, Sydney and Poland, and we support hundreds of PhD students in Google’s Fellowship Program.
As a global team, we continue to expand our footprint beyond our major hubs. Having solidified our research investment and innovation in Africa (Accra and Nairobi) and our presence in Australia, we are now preparing to inaugurate a new Google Research hub in Singapore in 2026.
We share our work through publications, conferences, academic talks, benchmarks, datasets and open-source releases. We’ve sponsored and hosted workshops at conferences, most recently at NeurIPS. We recently introduced an experimental program that provided automated feedback to scientists before they submit their conference papers for peer review, helping them to rigorously verify their work and accelerate research workflows. Plus, we launched Google Research Featured Notebooks in collaboration with NotebookLM, to make research more accessible to a broader community.
AI as an amplifier of human ingenuity
This is a golden age for research. Never before have technical breakthroughs and scientific progress so quickly materialized into impactful, real-world solutions, which, in turn, bring to the fore new data and questions that inspire new avenues of foundational research. This magic cycle is accelerating significantly, propelled by more powerful models, new agentic tools that support scientific discovery, and open platforms and tools.
Together with our Google colleagues and partners, we’re advancing research and technologies that aim to be helpful in diverse areas. Our research, grounded in a rigorous dedication to safety and trust, serves to unlock human potential — whether that’s to help a scientist accelerate their research, or a student learn more effectively and master new concepts, or to empower a doctor, developer or teacher.
It is truly an exciting time to be in research. We’re able to leverage the full stack of Google AI infrastructure, models, platforms, and world-class talent, and contribute to products used by billions. We will keep building on our legacy, asking the biggest questions of today, and aiming to enable the solutions of tomorrow. We’ll keep advancing AI in a bold and responsible way, for the benefit of society, to help enhance human capacity and make AI an amplifier of human ingenuity.
Acknowledgements
With thanks to everyone in Google Research, and many collaborators, who have contributed to this blog and the work represented here.