Machine intelligence
Google is at the forefront of innovation in Machine Intelligence, with active research exploring virtually all aspects of machine learning, including deep learning and more classical algorithms. Exploring theory as well as application, much of our work on language, speech, translation, visual processing, ranking and prediction relies on Machine Intelligence. In all of those tasks and many others, we gather large volumes of direct or indirect evidence of relationships of interest, applying learning algorithms to understand and generalize.
Machine Intelligence at Google raises deep scientific and engineering challenges, allowing us to contribute to the broader academic research community through technical talks and publications in major conferences and journals. Contrary to much of current theory and practice, the statistics of the data we observe shifts rapidly, the features of interest change as well, and the volume of data often requires enormous computation capacity. When learning systems are placed at the core of interactive services in a fast changing and sometimes adversarial environment, combinations of techniques including deep learning and statistical models need to be combined with ideas from control and game theory.
Recent Publications
Google Earth AI Geospatial Reasoning Agent: Capabilities and a Systematic Evaluation
Aliza Hoffman
Gal Moshitch
Aaron Bell
Roy Lee
William Cai
Adam Boulanger
Mark Young
Utkarsh Patel
Ruth Alcantara
Zachary Luna
Sydney Lin
Ollie Guinan
Christopher Haire
Lily Mihalkova
Luke Barrington
Tim Thelin
Gil Fidel
Ronald Ho
Samapriya Roy
Shravya Shetty
Mandar Sharma
Tyler McHenry
Joydeep Paul
Aviv Slobodkin
Jamie McPike
Noam Slonim
Per Bjornsson
Reuven Sayag
Allison Berman
Niv Efron
Gia Jung
arXiv (2026)
Preview abstract
Google Earth AI is a family of geospatial AI models, together with an agent that orchestrates them. We present this agent, Google Earth AI's Geospatial Reasoning agent (GRA), a Gemini-powered research system that combines geospatial data sources, foundation models, and tools to answer questions about our planet. GRA organizes these into modular domain skills and pairs them with upfront geospatial planning to decompose complex queries into multi-stage workflows. Evaluating such systems is hard: real-world problems admit multiple valid workflows, require data discovery across large catalogs, and rely on evolving ground truth. We introduce two benchmarks and release their development splits. GeoExpertQueries poses complex analytical problems grounded in practitioner use cases and refined by geospatial experts; each is paired with a consensus golden workflow from three independent experts and scored by a human-validated, tool-agnostic autorater. GeoKTools consists of templated queries with deterministic reference solvers that systematically vary the number of required tool types. We also use two external benchmarks, CBGB and UnivEARTH; the latter we find to be saturated. To isolate the contribution of GRA's tools and its agentic orchestration, we compare four configurations, each with Gemini 3.1 Pro and 3.8 Flash: a standalone LLM, the LLM with basic tools, the LLM with GRA's skills and tools but without its upfront planning and execution guidance, and GRA itself. With at most basic tools, an advanced model recovers much of the answer on templated queries but satisfies only a small fraction of the expert workflows on open-ended problems; GRA performs well on both. The largest step comes from GRA's skills and tools; on open-ended problems, its planning and execution guidance add a further gain with Gemini 3.8 Flash, letting Flash match or exceed Pro at lower cost.
View details
ToolGrad: Efficient Tool-use Dataset Generation with Textual "Gradients"
Kohei Uehara
Haoyu Zhang
Lin Gu
Zheng Xu
Tatsuya Harada
ACL 2026 (2026)
Preview abstract
Prior work synthesizes tool-use LLM datasets by first generating a user query, followed by complex tool-use annotations like depth-first search (DFS). This leads to inevitable annotation failures and low efficiency in data generation. We introduce ToolGrad, an agentic framework that inverts this paradigm. ToolGrad first constructs valid tool-use chains through an iterative process guided by textual "gradients", and then synthesizes corresponding user queries. This "answer-first" approach led to ToolGrad-500, a dataset generated with more complex tool use, lower cost, and almost 100% pass rate. Experiments show that ToolGrad models outperform those trained on expensive baseline datasets and proprietary LLMs.
View details
Assessing Global Flood Risk from Atmospheric Rivers through Physically Guided Machine Learning
Assaf Shmuel
Oleg Zlydenko
Martin Gauch
Colin Price
Scientific Reports (2026)
Preview abstract
Atmospheric rivers (ARs) are narrow corridors of concentrated moisture transport that play a crucial role in the global water cycle, delivering both beneficial rainfall and severe floods. Here, we develop a physically guided, explainable machine learning framework to predict flood occurrence during AR conditions worldwide by integrating AR characteristics with meteorological and topographic variables. The best model achieves a receiver operating characteristic area under the curve (ROC AUC) of 0.94, outperforming a logistic regression baseline at 0.81. Despite their limited footprint, we find that one third of large midlatitude floods occur under AR conditions, reflecting their disproportionate role in global flood risk. SHAP analysis highlights integrated vapor transport, precipitation, and elevation as dominant predictors. We show the non-linear amplification of flood risk under combined conditions of high soil moisture and persistent AR activity, underscoring the importance of antecedent wetness in modulating flood risk. Using consistent reanalysis inputs, we find that model-estimated high-flood-risk AR conditions increased globally by over 10% from 1980 to 2020 and shifted poleward. Validation on an independent satellite-based flood database shows comparable skill. We estimate that roughly 90% of the global population lives in regions that experience at least one AR annually, underscoring the broad societal relevance of AR dynamics. These findings highlight the value of physically guided machine learning for mapping and monitoring AR-related flood risk globally, offering actionable insights for preparedness and climate adaptation.
View details
Approaching the Privacy-Enhancing Technologies Adoption from a Requirements Engineering Perspective
Oleksandr Kosenkov
Danica Vranjanin
Abhinava Singh
Vadym Honcharenko
Volodymyr Spirin
2026
Preview abstract
In recent decades, Privacy Enhancing Technologies (PETs) have been gaining attention as a means to fulfill regulatory and user privacy requirements when processing personal data in software systems. Despite broad attention in research, support by regulators, contributions and adoption by large technological companies (like Google or Microsoft), and interest from software developers, adoption of PETs in practice is still challenging and lagging. For many years, existing research points to the same typical challenges to the adoption of PETs, such as complexity or absence of training; however, these challenges remain unresolved in practice despite the recent efforts.
In this industrial challenge paper, we take a deeper, practical requirements engineering-driven perspective on the challenges to PETs adoption across different stakeholders (namely, PETs developers and integrators, and PET adopters and users) and different disciplines (engineering, law, and business).
We suggest that the adoption of PETs should be facilitated with an in-depth understanding of engineering, business, and legal challenges to PETs, such as PETs' impact on software architecture, their business impact, and contribution to compliance. We suggest that requirements engineering research can play a crucial role in better exploring and modeling such viewpoints. In practice, requirements engineering could facilitate the coordination of the involved viewpoints to enable the adoption of existing and the development of new PETs.
View details
Towards conversational artificial intelligence for disease management
Khaled Saab
David Stutz
Yong Cheng
Kavita Kulkarni
Sara Mahdavi
Joelle Barral
James Manyika
Ryutaro Tanno
Adam Rodman
Nature (2026)
Preview abstract
Although large language models have shown promise in diagnostic dialogue, their capabilities for effective management reasoning, including disease progression, therapeutic response and safe medication prescription, have remained underexplored. We have advanced the previously demonstrated diagnostic capabilities of the Articulate Medical Intelligence Explorer (AMIE) using a new large-language-model-based agentic system optimized for multivisit clinical management and dialogue. To ground the reasoning of AMIE in authoritative clinical knowledge, we leveraged the long-context capabilities of Gemini, combining in-context retrieval with structured reasoning to align its output with up-to-date clinical practice guidelines and drug formularies. In a randomized, blinded virtual Objective Structured Clinical Examination study, AMIE was compared to 21 primary care physicians (PCPs) across 100 multivisit case scenarios designed to reflect the guidance of the UK National Institute for Health and Care Excellence and BMJ Best Practice guidelines. AMIE was non-inferior to PCPs in management reasoning, as assessed by specialists, and scored better both with respect to preciseness of treatment and investigation, and in terms of its alignment with and grounding in clinical guidelines. To benchmark medication reasoning, we developed RxQA, a multiple-choice question benchmark that was derived from two national drug formularies (from the USA and UK) and validated by board-certified pharmacists. Although AMIE and PCPs both benefited from the ability to access external drug information, AMIE outperformed PCPs on higher-difficulty questions. Although further research will be needed before real-world translation of AMIE, its strong performance across evaluations marks a significant step towards use of conversational artificial intelligence as a tool in disease management.
View details
Preview abstract
We introduce AMS (Activation-based Model Scanner), a tool that detects modifications to safety training in language models by measuring the geometric structure of safety-relevant concepts in activation space. Safety training creates measurable separation between harmful and benign content classes; certain safety modifications collapse or rotate this structure, while others leave it intact. We validate AMS across 14 model configurations spanning 4 architecture families (Llama, Gemma, Qwen, Mistral) and four safety-modification categories (instruction-tuned, base, abliterated, uncensored fine-tunes). Leave-one-out cross-validation of thresholds achieves 71% accuracy (10/14); bootstrap 95% confidence intervals on σ point estimates have median width 3.4σ and a substantial fraction of cells cross the PASS threshold under resampling. We further measure behavioral compliance on 20 stratified JailbreakBench prompts per model and find that σ on the harmful-content concept predicts compliance with Pearson r=−0.546 ( p=0.043 ); the rank-order Spearman correlation is weaker ( ρ=−0.423 , p=0.13 ). The structural signal predicts behavior directionally but with meaningful noise. Mechanistic analysis identifies a four-class taxonomy of safety-training modifications distinguished by activation-space signature: 1) training removal collapses cluster separation (e.g., base models, Dolphin variants: 0.5– 1.4σ ); 2) weight-orthogonalization-style abliteration both collapses separation and rotates the refusal direction (Llama-3.1-abliterated: σ=3.33 , direction cos sim 0.30); 3) rotation-without-collapse abliteration preserves cluster separation while rotating the refusal direction (Gemma-2-9b-abliterated: σ=4.54 , direction cos sim 0.84); and 4) behavioral fine-tuning that preserves both magnitude and direction (DarkIdol-1.2-Uncensored: σ=5.45 , direction preserved, 97% behavioral compliance). 1) and 2) AMS’s Tier 1 σ -threshold detects classes; 3) Tier 2 direction-similarity verification detects class; and 4) Class is undetectable by activation-only probing and represents a documented failure mode of the approach. We discuss threshold calibration, limitations of single-run measurement, and the open problem of detecting behavioral-only safety modifications.
View details
Some of our teams
×
![]()