Google Research’s cover photo
Google Research

Google Research

Technology, Information and Internet

Impossible? Let's see.

About us

From conducting fundamental research to influencing product development, our research teams have the opportunity to impact technology used by billions of people every day. We aspire to make discoveries that impact everyone, and sharing our research and tools to fuel progress in the field is fundamental to our approach.

Website
https://research.google/
Industry
Technology, Information and Internet
Company size
1,001-5,000 employees

Updates

  • Google Research reposted this

    Sequencing the genomes of millions of species sounds like a moonshot - because it is. ✨  To turn this vision into reality, at Google Research we are helping The Rockefeller University and the Earth BioGenome Project achieve their goal of sequencing all known species on Earth, from mammals and fish to plants and fungi. Today, we are sharing how Google’s AI tools have helped preserve the genetic information of 13 new endangered species—including the African penguin and the Cotton-top tamarin—and we are just getting started. Key AI advancements driving this work: ✨ DeepConsensus: Instrumental in removing errors from sequencing instruments to produce high-fidelity data. ✨ DeepPolisher: Our latest 2025 research which corrects remaining errors in genome assembly to reach the 99.999%+ accuracy required for comprehensive gene prediction. ✨ DeepVariant: A deep learning tool used by scientists to analyze the genomes of every living kākāpō in New Zealand, enabling a breeding plan that is pulling the species back from the brink of extinction. To further this mission, Google.org recently named The Rockefeller University as a recipient of the AI for Science fund to expand these efforts to 150 more species, all to be openly released to the scientific community and public. This project isn’t just cataloging life on Earth; it’s providing the insights necessary to prevent further loss of our planet’s biological legacy. It represents a key pillar of our research strategy: applying AI to solve "impossible" data challenges. For more than a decade, we have been building technology to accelerate scientific discovery, and genomics is a critical frontier. Read the full details on the Keyword blog: https://lnkd.in/djq7pwPm

  • Google’s AI tools are helping scientists sequence the DNA of every species on earth — and that's at least 1.5 million known species. Ten years of AI innovation for genomics at Google Research are contributing to this unprecedented scientific effort, resulting in a comprehensive catalog of life on earth. Read more about how we are we’re helping preserve the genetic code of endangered species with AI: https://goo.gle/3OiEN65 Learn more about our genomics tools here: https://lnkd.in/gbV6Qmat

  • A common heuristic in LLM agent design—"more agents is better"—might be wrong. Across 180 configurations, we find multi-agent coordination is task-contingent: +81% on parallelizable tasks (finance), but -70% on sequential ones (planning). Architecture-task alignment matters more than agent count.

    • Task-specific performance showing that multi-agent coordination yields substantial gains on parallelizable tasks like Finance-Agent (+81%) while degrading performance on sequential tasks like PlanCraft (-70%).
  • View organization page for Phasecraft

    9,038 followers

    Yesterday, we co-hosted a roundtable with Google Quantum AI at Google’s Washington D.C. office to discuss where quantum applications are headed, bottlenecks and areas for increased investment, and what’s required for near-term utility. Together with leaders from across the quantum computing industry and government, we explored: - How advances in quantum algorithms and hardware are converging toward practical capability, and soon - The role of public policy in accelerating meaningful deployment - Where quantum computing can begin to make a difference in energy, critical minerals, materials, pharma, and more - Why cooperation is critical to scaling applications Application and algorithmic development are essential to the future of quantum computing, and collaboration across science, industry, and government is necessary to unlock meaningful results in the near term. Thanks to our partners at Google and everyone who joined the conversation.

    • No alternative text description for this image
    • No alternative text description for this image
    • No alternative text description for this image
  • Introducing Greedy Independent Set Thresholding (GIST), a novel algorithm designed to select high-quality data subsets that maximize both data diversity and utility. As ML datasets grow in complexity, subset selection is critical for cost-effective training. GIST addresses this by finding a "sweet spot" between non-redundant data (diversity) and highly informative data (utility). Key highlights of this research include: Provable Guarantees: GIST is the first algorithm to provide a strong mathematical safety net, guaranteed to find a subset value at least half that of the absolute optimal solution. Superior Performance: In ImageNet benchmarks, GIST-enhanced strategies consistently outperformed state-of-the-art methods in single-shot data downsampling. Scalability: The algorithm's selection step is incredibly fast, making it practical for integration into pipelines with billions of data points. Presented at NeurIPS 2025, this work establishes a foundation for the next generation of scalable AI systems. Learn more: https://goo.gle/49R8XVe

    • A conceptual diagram of multiple overlapping colorful circles representing data clusters and diverse point selection.
  • The XPRIZE Quantum Applications competition is looking for teams turning quantum ideas into practical applications. Phase I finalists were selected based on real-world feasibility, rigorous benchmarks, and technical innovation. Curious what it takes to qualify and what’s coming next in Phase II? Take a look → https://goo.gle/3NAEYti

  • Introducing a novel decomposition approach for improved intent extraction using small multimodal models. Understanding user intent from UI interactions on mobile devices often requires large, server-side models. Our researchers have developed a two-stage workflow that first summarizes individual interactions and then extracts intent from the sequence. This method allows small models, such as Gemini 1.5 Flash 8B, to match the performance of much larger models like Gemini 1.5 Pro. Presented at #EMNLP2025, this research highlights how decomposing complex tasks can make on-device AI more tractable, private, and efficient. Learn more: https://goo.gle/4jSej7j

    • A diagram titled that illustrates a user's intent extraction workflow. The user selects from a grid of travel options. Below, text labels describe the screen context, the specific user action, and a speculation on the user's travel goals.
  • Introducing our latest research on smartwatch-based walking metrics estimation. Historically, capturing precise gait metrics like double support time or swing time required specialized lab equipment. Our team has developed a multi-head temporal convolutional network that allows a single smartwatch to estimate these vital health biomarkers with high reliability. Through a study with 246 participants, we verified that smartwatch-based estimates are comparable to smartphone-based methods. This work is a significant step toward making longitudinal gait monitoring accessible for applications such as early disease detection, fall risk prevention, and personalized rehabilitation planning. Learn more: https://goo.gle/4pEQBfW

    • Gait parameter accuracy reflecting the mean absolute percentage error (MAPE) for Pixel Smartwatch (Watch) and Pixel Smartphone (Phone) (N=246 participants). Boxes indicate the interquartile range (Q1–Q3), whiskers show 1st–99th percentiles.

Affiliated pages

Similar pages