google·Jul 30, 2026·3 min readCurated summary

Science One Framework: A verifiable autonomous research framework via Chain-of-Evidence

The Science One Framework addresses a central weakness in autonomous AI research: polished papers can contain fabricated citations, unreproducible results, or methods that do not match the underlying code. Its Chain-of-Evidence (CoE) framework requires every claim to be linked to concrete evidence and introduces CoE Audit to test those links automatically. In evaluations, Science One produced fully verifiable papers while matching or exceeding baseline research agents and human performance on several benchmarks. ## Chain-of-Evidence for Verifiable Research - CoE defines trustworthy research artifacts through two requirements: - **Completeness:** Every claim has a recorded evidence chain. - **Correctness:** The evidence genuinely supports the claim. - Claims may include: - Bibliographic references - Reported scores - Method descriptions - Conclusions - Evidence can include peer-reviewed papers, experiment logs, executed code, or result tables. - Hallucinated citations, unreproducible scores, and discrepancies between described and implemented methods are treated as broken evidence chains. ## The Science One Framework The framework builds evidence into the research process instead of attempting to verify a paper after it has been written. - **Problem Investigator** - Uses the Semantic Scholar API to construct a citation graph. - Reads up to 100 full-text PDFs per topic. - Produces a structured research brief. - Restricts final-paper references to sources retrieved through the grounded API, avoiding citations generated from model memory. - **Discovery Engine** - Explores ideas through parallel explore-exploit branches. - Each isolated cycle includes a Solver agent and a task-specific evaluator. - High-performing solutions are iteratively refined. - Raw evaluator outputs are preserved in strict, read-only records. - **Paper Writer and Claim Verifier** - Creates a structured inventory of factual claims. - Attaches inline evidence tags linking claims to workspace artifacts. - Checks each claim against its declared source. - Rewrites unsupported claims conservatively rather than allowing them to exceed the evidence. ## CoE Audit Integrity Checks CoE Audit is an automated, post-hoc forensic review of a paper, solution, code, and references. - **Score verification:** Re-runs the submitted code independently and compares the result with the paper’s reported score. - **Specification violation:** Checks whether the code solves the intended task without exploiting the evaluator or accessing ground-truth answers. - **Reference verification:** Validates every bibliography entry against academic APIs. - **Method-code alignment:** Compares the paper’s method description with the actual implementation using LLM-based judges. ## Evaluation Results - The audit evaluated 75 papers across five systems-optimization tasks: Prism, Cloudcast, EPLB, LLM-SQL, and transaction scheduling. - Science One led the evaluated systems on all four integrity checks. - It had: - Zero phantom references - Perfect score verification - The strongest method-code alignment - Baseline systems hallucinated up to 21% of references and sometimes described advanced algorithms that were implemented as simple deterministic heuristics. - Strong verification did not reduce performance: - Science One matched or exceeded human experts on all five ADRS tasks. - It achieved the best overall result on Cloudcast and EPLB. - On additional MLE-Bench and Parameter Golf evaluations, the framework also demonstrated competitive performance, including two Gold Medals across five difficult Kaggle competitions. The main recommendation is to design autonomous research systems around evidence generation from the beginning. Grounded retrieval, immutable experiment records, claim-level verification, and independent auditing can substantially improve reliability without necessarily sacrificing research performance.

Read original(opens in new tab)

Machine LearningLarge Language ModelsAutonomous AgentsSemantic Scholar Api+3

google·Jul 22, 2026·3 min readCurated summary

Towards a quantum computer that learns from its errors

Quantum computers require constant recalibration because analog control signals drift during computation. Google Quantum AI combined reinforcement learning (RL) with quantum error correction (QEC), allowing a system to learn from detected errors and adjust thousands of control parameters while computation continues. Tests on the Willow processor showed improved logical stability, suggesting this approach could support much longer quantum computations. ## The Challenge of Quantum Errors - Quantum systems are highly sensitive to drift in signal frequencies, amplitudes, and phases. - Conventional recalibration requires stopping the entire computation, limiting algorithms that may need to run for days or months. - QEC uses many physical qubits to form logical qubits and converts analog noise into binary error-detection events. - These events indicate that an error occurred within a spacetime region but do not identify its precise cause or location. - Decoders such as AlphaQubit and Tesseract infer corrections, but they do not explain whether errors arose from environmental decoherence or preventable calibration drift. ## Moving Beyond Physics-Based Calibration - Traditional calibration depends on manually designed physical models. - Such models can reach performance limits when hardware behavior involves complex, poorly understood interactions. - Google argues that quantum control may benefit from the same shift toward data-driven learning seen in computer vision, robotics, and protein-folding research. - As quantum hardware improves, remaining errors increasingly reflect subtle phenomena that are difficult to model analytically. ## Using Error Detection as a Learning Signal - An RL agent experiments with control strategies and improves based on the resulting error data. - QEC detection events serve two purposes: - Decoders use them to infer logical corrections. - The RL system uses them to identify drift and refine control parameters. - This enables continuous calibration without interrupting the quantum computation. - The approach can steer thousands of analog control parameters dynamically. ## Results on the Willow Processor - Researchers deliberately introduced control-parameter drift into Google’s Willow superconducting processor. - RL steering improved the logical stability of the error-correcting code by 3.5 times. - After expert, human-guided calibration, RL fine-tuning reduced the logical error rate by an additional 20%. - Combined improvements produced fewer than one logical error per 1,000 surface-code correction cycles and fewer than one per 100 color-code cycles. - The processor therefore operated as a more reliable quantum memory for longer periods. ## Scaling to Larger Systems - Simulations included hundreds of qubits and tens of thousands of control parameters. - The RL agent reduced initially high physical error rates by learning better control settings. - QEC suppressed the logical error rate exponentially as the number of physical qubits increased. - The simulations indicated that the number of RL training iterations needed to reduce physical errors did not depend on system size, supporting potential scalability. The results suggest that future quantum computers could use QEC not only to correct errors but also to learn their causes and continuously adapt to hardware drift. RL-based calibration could reduce dependence on manual tuning and help make long-running, fault-tolerant quantum computation practical.

Read original(opens in new tab)

Machine LearningQuantum ComputingReinforcement LearningSuperconducting Qubits+4

google·Jun 30, 2026·3 min readCurated summary

Expanding our Heat Resilience data to 50+ global cities

Google Research has expanded its building-level rooftop reflectivity dataset from 14 pilot cities to more than 50 cities across nine countries. By combining Sentinel-2 satellite data with 30-centimeter commercial imagery and machine learning, the project identifies roofs where reflective “cool roof” interventions could reduce urban heat. The data is publicly available through a new Google Earth Engine app for planners and researchers. ## Why Rooftop Reflectivity Matters - Extreme heat contributes to approximately 500,000 deaths annually. - Urban heat islands make cities warm faster than the global average. - Dark roofs, pavements, and limited vegetation increase heat absorption. - Reflective roofs reduce the solar energy absorbed by buildings and can lower local surface temperatures. - Earlier pilot data from 14 cities supported cool-roof ordinances and broader climate adaptation plans. ## Building-Level Albedo Mapping - Sentinel-2 provides global albedo data but at 10-meter resolution, which cannot reliably distinguish individual rooftops. - Google Research fused Sentinel-2’s spectral and global coverage with 30-centimeter Airbus Pléiades Neo imagery. - Machine learning and radiometric calibration reconstruct detailed reflectance profiles at the urban-pixel level. - Validation against airborne hyperspectral measurements in Boulder, Colorado, produced an RMSE of 0.04. - The resulting maps allow planners to prioritize large, low-reflectivity buildings for cool-roof retrofits. - The modeling suggests targeted interventions could reduce extreme urban heat by up to 0.5°C globally. ## Heat Resilience Earth Engine App - Displays building-level albedo using rooftop centroids to highlight low-reflectivity surfaces. - Provides baseline analyses and supports monitoring changes over time. - Allows users to download high-resolution data for local studies and policy development. - Offers a nested view that moves from census-tract summaries to individual buildings. ## Expanded Global Coverage - The dataset now covers more than 50 cities in nine countries. - Newly included urban areas span Europe, Brazil, and the United States. - Examples include London, Athens, Barcelona, Rio de Janeiro, São Paulo, Los Angeles, Austin, and New York City. - The open dataset is intended to help municipalities accelerate reflective-surface programs. ## Access and Collaboration - The interactive app and datasets are publicly available through Google’s Heat Resilience site. - The methodology is described in the Nature Communications paper “Estimating high-resolution albedo for urban applications.” - The work was developed by Google Research in collaboration with the World Resources Institute. Cities can use the app to identify the buildings and neighborhoods where cool-roof investments are likely to have the greatest heat-reduction benefits.

Read original(opens in new tab)

Machine LearningGeospatial DataSatellite ImageryEarth Engine+3

google·Jun 25, 2026·3 min readCurated summary

Optimizing cloud economics with linear elastic caching

Linear elastic caching treats cache memory as a variable cost rather than a fixed allocation. It dynamically adjusts how long pages remain in memory by balancing ongoing memory expense against the cost of fetching evicted data again, using the ski rental problem as its theoretical foundation. Experiments in Spanner and public cache traces show meaningful cost reductions with only modest increases in misses. ## Fixed-Size Cache Limitations - Traditional caches allocate a fixed amount of RAM and use policies such as LRU when space runs out. - Undersizing the cache causes excessive disk or storage access and poor performance. - Oversizing it wastes money during periods of low demand; some serverless providers charge up to $3 per day for 1 GiB of memory. - Fixed sizing therefore creates a “Goldilocks” problem as workloads fluctuate. ## Ski Rental Model for Cache Eviction - Each cached page presents two choices: - **Rent:** Keep it in RAM and continuously pay for its memory footprint. - **Buy the miss:** Evict it and risk a latency and I/O penalty if it is requested again. - A ski rental algorithm assigns each page a time-to-live (TTL). - If the page is not accessed before its TTL expires, it is evicted. - If the cache becomes physically full, a conventional policy such as LRU handles capacity pressure. - The researchers prove that eviction policy and rental duration can be optimized separately, simplifying implementation. - Unlike worst-case break-even or randomized ski rental strategies, lightweight machine learning can exploit predictable workload patterns. ## Lightweight TTL Prediction - In Spanner, each page receives a TTL based on: - Page size - Cost of a cache miss - Type of database operation - Observed access behavior - A shallow decision tree was chosen because Spanner processes billions of requests per second. - The model can be translated into a few lines of interpretable C++ code. - Its cost-aware decisions allow extra misses mainly for data that is inexpensive to retrieve. ## Spanner Production Results - Compared with a standard fixed-size cache: - Memory usage fell by **15.5%**. - Cache misses increased by only **5.5%**. - Total cost of ownership fell by approximately **5%**. - The additional misses increased actual I/O costs by only **0.5%**, because they were concentrated on cheap-to-fetch data. - The policy was deployed on production Spanner servers and evaluated over several months. ## Public Trace Evaluation - The approach was tested on public industry cache traces using GDSF as the fixed-size baseline. - GDSF generalizes LRU to account for pages with different sizes. - Researchers evaluated four elastic-cache variants using: - Break-even or randomized ski rental policies - Learned or non-learned TTL selection - Because public traces lacked application-level features, learning used the first half of each trace to calculate the best TTL for individual pages. - Caches were warmed with one day of requests before performance measurement began. ## Overall Results - Elastic caching consistently produced lower total cost across diverse workloads. - Its advantage increased as memory became more expensive relative to cache misses. - At comparable cache sizes, elastic policies also achieved substantially lower miss rates than fixed-size approaches. Linear elastic caching is most useful when memory costs vary significantly or workloads are bursty and predictable. Dynamically assigning page TTLs offers a practical way to reduce memory spending while limiting performance impact, especially when the system can estimate the cost of each miss.

Read original(opens in new tab)

Machine LearningCachingC Plus PlusDatabase Systems+4

google·Jun 12, 2026·3 min readCurated summary

Research into how AI can help users understand skin conditions

Google Research examined how AI tools can help non-experts understand skin conditions and decide what to do next. In a large study, AI substantially improved people’s ability to identify possible conditions, but it did not reliably improve the accuracy of recommended next steps. The research therefore argues that dermatology AI should be designed around human decision-making, safety, and clear guidance—not diagnosis alone. ## Why Dermatology AI Needs Human-Centered Research - More than half of adults use the Internet for health information, and about one-third use AI. - People often lack the medical vocabulary needed to search effectively—for example, searching for “red dots on legs” instead of “palpable purpura.” - Google Research has developed dermatology AI models, validated their generalization, and released datasets such as SCIN. - Earlier research found that online tools can improve condition recognition without necessarily helping people choose appropriate next steps. - The researchers emphasize studying how people interpret and act on AI-generated information. ## Large-Scale Evaluation of an AI Information Tool - A JAMA Dermatology study involved 2,345 participants reviewing de-identified skin-condition cases with images and structured medical histories. - Participants were assigned to one of three groups: - **Standard-search control:** Used familiar text-based search tools. - **AI group:** Used a prototype showing 3–7 AI-predicted conditions, textbook images, and information about symptoms and treatments. - **“Wizard of Oz” control:** Used the same interface, but with dermatologist-provided differential diagnoses presented as if generated by AI. - The AI interface increased participants’ willingness to name a condition: - More than 62% attempted a diagnosis with AI. - Only 41% did so using standard search. - Accuracy also improved: - AI users correctly identified a matching condition about 23% of the time. - Standard-search users achieved 8%. - The “perfect-prediction” interface reached 36%, showing that even accurate candidate lists did not make users nearly perfect. - AI users reported greater confidence, satisfaction, and satisfaction with the time spent searching. ## Identifying a Condition Does Not Guarantee Safe Action - The prototype intentionally avoided prescribing actions or making individualized diagnoses. - Treatment information was dermatologist-written and based on the condition name, rather than the severity or details of the specific case. - Choosing the right next step—such as home care, routine care, or urgent evaluation—remained difficult. - Next-step accuracy improved only slightly in the “Wizard of Oz” group, from 60% in the standard-search control to 63.5%. - The standard AI group showed no statistically significant improvement. - AI users were slightly more likely than control participants to recommend a less urgent action than dermatologists would: 30% versus 27%. - These findings show that identifying possible conditions is insufficient without stronger safety-oriented guidance. ## Studying Real Users and Diverse Communities - The researchers also conducted a qualitative study, published at ACM CHI, to examine how people use AI for their own active skin concerns. - The project partnered with Stanford’s Healthcare AI Applied Research Team and the Santa Clara Family Health Plan. - The community included many Medi-Cal users who rely on a healthcare safety net. - Researchers aimed to gather richer feedback than survey-based studies provide by observing real-world use. - Because participants spoke four primary languages, the application was translated into those languages, with multilingual volunteers or staff available to support communication. AI can make dermatology information easier to find and improve recognition of possible conditions, but it should not be treated as a substitute for professional judgment. Future tools should focus equally on urgency assessment, personalized context, uncertainty, and clear recommendations for when to seek medical care.

Read original(opens in new tab)

Machine LearningAIHealthcare AiDifferential Diagnosis+3

google·Jun 3, 2026·3 min readCurated summary

The next chapter in flood resilience: Open sourcing Google’s hydrology framework

Google Research has open-sourced the hydrology framework behind its Flood Hub river forecasts. The Python/PyTorch package lets researchers and national forecasting agencies train AI models with global and local data while retaining control over their information. Google argues that open access, local expertise, and interoperable tools can make advanced flood warnings more accurate, affordable, and widely deployable. ## The Open-Source Hydrology Framework - The framework is available on GitHub under an Apache 2.0 license. - It provides model architectures, training pipelines, documentation, and tutorials. - Users can train models with climate, soil, topography, land-cover, and weather data. - Historical river observations come from the open Caravan dataset, which agencies can extend with local measurements. - The package is built with PyTorch and is intended for both researchers and operational forecasters. ## Model Versions and Improvements - The release includes: - The original model used in Google’s 2024 benchmarking study. - An upgraded v2 model currently used for real-time global forecasts in Flood Hub. - The v2 model uses a multi-input ME-LSTM architecture. - Separate networks embed different meteorological products before combining them in an LSTM. - Inputs include GraphCast, ECMWF forecasts, NASA IMERG satellite rainfall estimates, and NOAA CPC precipitation data. - Benchmarking showed the newer model extends the reliable forecast horizon by: - Six days in gauged river basins. - One day in ungauged basins. ## Local Data and Operational Forecasting - Agencies can fine-tune models for specific watersheds using local observations and expert knowledge. - The approach supports the integration of Indigenous and Local Knowledge, which the World Meteorological Organization says is still rarely incorporated systematically. - Models are designed to be relatively inexpensive and easier to train than traditional conceptual hydrological systems. - Local organizations can preserve control over their data while adapting the models to regional conditions. ## Partnership with the Czech Hydrometeorological Institute - Google worked with CHMI to validate the model against locally calibrated traditional forecasting models. - CHMI created an adapter connecting the framework to Delft-FEWS, a widely used operational forecasting platform. - This integration demonstrates how machine-learning forecasts can fit into existing workflows used by government agencies, NGOs, and private organizations. - The partnership provides a practical model for other national hydrological services. ## Broader Flood-Resilience Goals - Open-source distribution could help resource-constrained regions access advanced forecasting without expensive infrastructure. - The framework is intended to support capacity building for early-warning systems worldwide. - Google presents the release as a way to let the global hydrology community reproduce, improve, and localize its research. National hydrological agencies and researchers should evaluate the open-source framework using their own watershed data, integrate it with existing forecasting systems, and validate its predictions against established local models before operational deployment.

Read original(opens in new tab)

Machine LearningPythonPytorchFlood Forecasting+3

google·May 28, 2026·3 min readCurated summary

A New Era of Innovation: Google Research at I/O 2026

Google’s I/O 2026 research announcements present AI as an “agentic” amplifier of human ingenuity, particularly in science and healthcare. New systems such as Gemini for Science, ERA, Co-Scientist, and Gemini Deep Think are designed to generate hypotheses, write and optimize code, evaluate evidence, and solve difficult research problems. Google also highlighted health-focused AI that supports users before, during, and after medical visits, while emphasizing collaboration, validation, and responsible deployment. ## AI-Driven Scientific Discovery - **Gemini for Science** is a suite of experimental tools built from Google Research and developed with Google Cloud, Google DeepMind, and Google Labs. - **Empirical Research Assistance (ERA)** acts as a code-optimizing research engine: - Proposes concepts and writes software. - Evaluates results against a defined scoring system. - Uses tree search to test thousands of code variants. - Has supported work in neuroscience, cosmology, respiratory-illness forecasting, and California runoff prediction. - **Co-Scientist** is a Gemini-based multi-agent collaborator that generates, evaluates, and refines hypotheses. - Researchers have applied it to antimicrobial resistance, plant immunity, and liver fibrosis. - **Computational Discovery**, combining ERA and AlphaEvolve, runs thousands of code variations in parallel to test scientific models and hypotheses more quickly. - **Hypothesis Generation** uses a multi-agent “idea tournament” to debate and rank research ideas, with clickable citations supporting claims. - **Literature Insights**, powered by NotebookLM, helps researchers synthesize large bodies of scientific literature. - **Science Skills** can automate specialist workflows such as structural bioinformatics and genomic analysis on agentic coding platforms. ## AI for Peer Review and Advanced Reasoning - Google is piloting the **Paper Assistant Tool (PAT)** for scientific peer review. - PAT has experimentally reviewed more than 10,000 papers for conferences including ICML, STOC, and NeurIPS. - Its feedback has helped authors identify theoretical gaps and design additional experiments. - **Gemini Deep Think** has been used with mathematicians, physicists, and computer scientists to address open problems involving network deadlocks, optimization, machine-learning behavior, auction theory, and cosmic-string singularities. ## Advancing Health with AI - Google’s health research focuses on supporting people throughout the full healthcare journey, from understanding symptoms and preparing for appointments to interpreting medical records. - Research contributions underpin the **Google Health app** and **Google Health Coach**, with the app beginning rollout to existing Fitbit users. - **Symptom AI** investigates how conversational AI can reason about information relevant to a person’s symptoms. - A Fitbit-based study included 13,917 participants. - In blind comparisons, clinicians preferred Symptom AI’s differential diagnoses roughly twice as often as those produced by other clinicians. - The **Plan for Care** pilot involved 1,779 participants preparing for doctor visits. - Compared with baseline systems, 15% more users felt prepared. - 13% more users felt confident they could make effective use of their appointment. - Google is also studying personal health large language models and the use of personal health record data to improve health guidance. Google’s announcements point toward research systems that actively experiment, collaborate, and reason rather than merely retrieve information. Their practical value will depend on continued scientific validation, clinician involvement, privacy protections, and careful expansion from experimental tools into real-world use.

Read original(opens in new tab)

Machine LearningGenerative AIGeminiMulti Agent Systems+2

google·May 19, 2026·2 min readCurated summary

Empirical Research Assistance (ERA): From Nature publication to catalyzing Computational Discovery

Empirical Research Assistance (ERA) is a Google AI system designed to help scientists develop expert-level computational models. Using Gemini, it searches literature, generates and evaluates code, and explores thousands of possible solutions through tree search. A Nature paper reports strong performance across scientific benchmarks, while new applications suggest ERA can accelerate research in health, climate, energy, and economics. ## How ERA Supports Scientific Coding - ERA starts with a scientific problem and a success metric. - It searches relevant research, combines methods, writes code, and iteratively tests and improves solutions. - Its tree-search process evaluates thousands of alternatives to optimize the resulting model. - Benchmarks in genomics, public health, satellite imagery, neuroscience, time-series forecasting, and mathematics showed expert-level performance. ## Applications to Open Scientific Problems - **Epidemiological forecasting** - Predicted U.S. hospital admissions up to four weeks ahead for flu, COVID-19, and RSV. - Forecasts ranked at or near the top of CDC leaderboards. - The techniques can potentially be adapted to other countries and diseases. - **California water-supply forecasting** - Produced seasonal runoff predictions for snow-fed river basins. - Delivered more accurate early forecasts than California’s official Bulletin 120 outlook. - Improved predictions could support water management and agriculture. - **Atmospheric carbon dioxide monitoring** - Combined geostationary weather-satellite data with other inputs to estimate CO₂ concentrations every 10 minutes across broad areas. - Captured urban emissions, plant-driven daytime absorption, and other atmospheric cycles. - Provides higher spatial and temporal coverage than measurements from satellites such as Orbiting Carbon Observatory-2. - **Solar-energy design** - Combined ERA with Google Antigravity to optimize three-dimensional solar-panel geometries. - Identified a 500-triangle volumetric fan design that could capture scattered radiation without backward shading. - **Retail forecasting** - Used economic indicators, Google Trends, historical patterns, and consumer sentiment. - Matched or exceeded commercial consensus forecasts and the Chicago Fed’s monthly retail forecast. ## Computational Discovery - Google is gradually opening access to Computational Discovery through a trusted tester program in Google Labs. - The system combines ERA with AlphaEvolve to support computational scientific investigation. - It complements other Gemini for Science experiments: - **Hypothesis Generation**, built with AI Co-Scientist, supports developing scientific hypotheses. - **Literature Insights** supports research and literature analysis. ERA’s demonstrated value lies in automating the labor-intensive cycle of designing, testing, and refining scientific software. Its expanding applications indicate that AI-assisted computational research could broaden access to advanced modeling while helping experts investigate complex scientific problems more quickly.

Read original(opens in new tab)

Machine LearningAIGeminiTime Series Forecasting+4

google·Apr 22, 2026·2 min readCurated summary

It's all about the angle: Your photos, re-composed

Google’s new Auto frame feature in Google Photos re-composes photos by changing the apparent camera viewpoint after capture. It combines 3D scene reconstruction with generative AI to adjust framing, reduce wide-angle distortion, and generate background areas that were outside the original frame. The result is a more natural perspective while preserving the visible content and subjects’ identities. ## Why Traditional Editing Falls Short - Cropping and zooming cannot change the original camera perspective or parallax. - They also cannot reveal areas hidden outside the captured frame. - This is especially limiting for portraits, selfies, and “almost perfect” photos with poor framing or lens distortion. ## Two-Stage 3D-Aware Editing - **3D scene and camera estimation** - A machine-learning model creates a 3D point map from the 2D image using monocular depth estimation. - Each pixel is associated with an estimated 3D surface point. - The system estimates the original focal length and reconstructs human faces and bodies carefully to reduce identity-distorting artifacts. - Semantic analysis detects subjects’ face positions and 3D orientations to determine better camera parameters. - **Generative completion and retouching** - Classical 3D rendering produces an image from the modified camera position, orientation, and focal length. - Moving the virtual camera exposes “holes” where the original image contained no information. - A latent diffusion model fills these missing regions and corrects the rendered result. - Classifier guidance with regional scaling helps preserve original content while allowing generation in newly revealed areas. ## Automatic Framing and Distortion Correction - The system automatically selects camera parameters for improved portrait composition. - It can adjust both camera pose and intrinsic properties such as focal length. - For wide-angle selfies, it detects perspective distortion that enlarges features close to the lens. - By effectively moving the virtual camera backward and adjusting the optics, it produces more natural facial proportions. ## Availability in Google Photos - The technology is now integrated into the Auto frame feature. - It applies to eligible photos containing people. - Users receive the re-composed image as a second Auto frame candidate. - The enhancement requires only a single action and automatically changes the viewpoint while retaining the original scene. The feature demonstrates how combining geometric 3D reconstruction with controlled generative inpainting can perform edits that conventional cropping and retouching cannot. For portraits and selfies, Auto frame offers a practical way to improve composition and perspective after the moment has passed.

Read original(opens in new tab)

Machine LearningGenerative AI3d RenderingImage Editing+4

google·Mar 31, 2026·3 min readCurated summary

Building better AI benchmarks: How many raters are enough?

Human disagreement makes AI benchmarks difficult to reproduce, yet evaluations often use only one to five raters per item and reduce their responses to a majority vote. The study introduces an `(N, K)` framework—balancing the number of items (`N`) against raters per item (`K`)—to determine how annotation budgets should be allocated. It concludes that the best balance depends on the evaluation goal: broad sampling for majority accuracy, but deeper rating for capturing nuanced human opinions. ## The Breadth-versus-Depth Trade-off - The “forest” strategy rates many items with few raters per item. - The “tree” strategy rates fewer items with many raters per item. - Historically, AI benchmarks have favored the forest approach, typically using one to five raters per example. - This approach can miss both the overall distribution of opinions and meaningful disagreement among raters. ## Simulating Annotation Budgets - The researchers built a simulator using real-world subjective datasets, including toxicity, hate speech, safety, offensiveness, and job-related tweet classification. - They varied: - **Scale (`N`)**: 100 to 50,000 total items. - **Crowd (`K`)**: 1 to 500 raters per item. - Thousands of configurations were tested for statistical reliability, including whether model comparisons reached significance at `p < 0.05`. - The simulator also examined messy conditions such as highly imbalanced categories and tasks with multiple labels. - The simulator has been released as open source. ## Why Three to Five Raters Are Often Insufficient - Low-rater evaluations may fail to represent natural human disagreement. - They provide too little depth to reveal nuanced opinions and too little breadth to establish a reliable overall picture. - In many settings, more than 10 raters per item are needed to produce results that reflect the variation in human judgments. - More ratings per item can make model comparisons more statistically reliable. ## The Evaluation Metric Determines the Optimal Strategy - **Majority-vote accuracy** - If the goal is to determine whether a model agrees with the majority of people, rating more items is generally more effective. - This favors the forest strategy. - **Opinion range and nuance** - If the evaluation must distinguish between responses such as “yes,” “maybe,” and “no,” more raters per item are essential. - This favors the tree strategy because only repeated ratings reveal the full distribution of human opinions. - There is no universally optimal number of items or raters; the correct allocation depends on what the benchmark is intended to measure. ## Reproducibility Without Unlimited Budgets - An appropriately chosen item-to-rater ratio can produce highly reproducible results with roughly 1,000 total annotations in some settings. - Spending more money does not guarantee reliability if the budget is distributed poorly. - The study’s framework is intended to help benchmark designers choose the allocation that best fits their metric and data characteristics. ## Moving Beyond a Single Ground Truth - Many AI evaluations assume that every example has one objectively correct label. - This assumption becomes increasingly problematic for subjective tasks involving toxicity, harmful intent, ethics, safety, or social interaction. - Preserving disagreement instead of collapsing it into a plurality label can make benchmarks more representative of real human judgment. - The authors argue that understanding disagreement is as important as measuring consensus. Benchmark designers should first decide whether they need majority accuracy or a detailed picture of human opinion, then allocate ratings accordingly. In subjective evaluations, using substantially more than five raters per item may be necessary for reliable and reproducible conclusions.

Read original(opens in new tab)

Machine LearningOpen SourceData AnnotationReproducibility+3

google·Mar 17, 2026·3 min readCurated summary

Improving breast cancer screening workflows with machine learning

Google Research’s AIMS studies evaluated whether machine learning could support the UK’s mammography double-reading workflow. Across five NHS screening services, the AI system improved cancer detection sensitivity without reducing specificity, detected some cancers missed by human readers, and processed cases far faster. The studies also showed that safe deployment requires local calibration, monitoring for distribution shifts, and evaluation of how clinicians interact with AI results. ## NHS Screening Challenges - The UK NHS uses two human readers for each mammogram, with arbitration when their assessments require review. - A projected shortage of clinical radiologists—currently around 30% and expected to reach 40% by 2028—threatens the sustainability of this model. - AI could help increase detection while reducing pressure on radiology services. ## Study 1: Standalone Performance - The retrospective evaluation included mammograms from approximately 116,000 women screened across five NHS services. - The services represented three different double-reading and arbitration workflows. - AI thresholds were calibrated separately for each service to account for local populations and procedures. - Performance was measured against the original first reader using a 39-month follow-up period, including interval and subsequent-round cancers. - Researchers also assessed: - Comparisons with second and consensus readers - Lesion-level localization - Performance across demographic groups ## Study 1: Results - Cancer detection increased from 7.54 to 9.33 cases per 1,000 women. - The AI system achieved significantly higher sensitivity than the original first reader without compromising specificity. - It detected 25% of interval cancers missed by the original double-reading process. - Performance was especially strong for invasive cancers and women attending their first screening. - The study found no notable systematic disparities by age, ethnicity, breast density, or socioeconomic status. ## Prospective Technical Deployment - The system was deployed non-interventionally at 12 sites across two London screening services. - It processed 9,266 cases over roughly two months per service. - Mammograms were pseudonymized and sent to a secure Google Cloud-based system. - Median AI processing time was 17.7 minutes, compared with more than two days for the first human read. - The deployment detected a distribution shift between historical training data and current clinical data. - Researchers adjusted operating points during deployment to maintain safe and appropriate recall rates for local workflows. ## Study 2: AI in the Double-Reading Workflow - The second study examined how human readers performed when using AI as part of arbitration, rather than evaluating AI in isolation. - Twenty-two readers reviewed thousands of cases using real screening-service rules. - Two workflows were compared: - **Standard care:** decisions from the historical first and second human readers - **AI-enabled care:** the historical first-reader decision paired with the AI decision - This design aimed to assess the practical effects of replacing the second human read with an AI reader. The findings support AI as a potential second reader in breast cancer screening, but broader prospective clinical validation is still needed. Successful adoption should include phased deployment, local calibration, continuous monitoring, and careful evaluation of human-AI decision-making.

Read original(opens in new tab)

Machine LearningAIDatabase DesignCloud Computing+4

google·Mar 12, 2026·3 min readCurated summary

Protecting cities with AI-driven flash flood forecasting

Google Research is expanding Flood Hub with urban flash flood forecasts that can provide up to 24 hours’ warning. The system addresses the lack of historical flood observations by using Gemini to extract verified events from public news reports, creating the Groundsource dataset for model training. Its global, lower-resolution approach aims to extend useful warnings to regions that lack expensive sensors and forecasting infrastructure, particularly in the Global South. ## The Need for Earlier Flash Flood Warnings - Flash floods cause roughly 85% of flood-related deaths worldwide and kill more than 5,000 people annually. - They often develop within six hours of intense rainfall, making rapid warnings essential. - Even 12 hours of warning can reduce flood damage by about 60%. - Early warning coverage remains highly unequal: fewer than half of developing countries have access to multi-hazard warning systems. - Flood Hub previously focused mainly on slower-moving riverine floods, covering more than 2 billion people across 150 countries. ## The Data Problem: “Invisible” Floods - River flood models can rely on stream gauges that record water levels and flow. - Flash floods may occur far from gauges, especially in cities where rainfall, impermeable surfaces, drainage, and terrain interact unpredictably. - Building detailed physical simulations globally would be computationally expensive. - Historical, precisely located flash flood records are also scarce, preventing conventional supervised machine learning. - Google’s Groundsource method uses Gemini to analyze public news reports, verify flood locations and times, and assemble a historical flash flood dataset. ## Scaling from Local Systems to Global Coverage - Local flash flood systems can be highly accurate using rain sensors, radar, water-level monitors, and flow measurements. - These systems are expensive to deploy and require location-specific calibration and engineering expertise. - Broader systems such as WMO’s FFGS, ERIC, and the U.S. NWS warning system depend on high-resolution maps, radar forecasts, and skilled hydrologists. - Those resources are often unavailable in the Global South. - Google’s model instead uses globally available products, including NASA IMERG, NOAA CPC, ECMWF’s IFS HRES forecasts, and Google DeepMind’s medium-range weather model. - Forecasts currently operate at a 20-by-20-kilometer resolution, constrained by the resolution of global data sources. ## The Urban Flash Flood Model - The model estimates whether a flash flood is likely in a given area during the next 24 hours. - It uses a recurrent neural network with a long short-term memory (LSTM) component to process meteorological time series. - Inputs also include static geographic and human-environment factors: - Urbanization density - Topography - Soil absorption rates - The initial rollout targets urban regions, where news coverage is denser and most of the world’s population lives. - It currently predicts impacts in areas with population densities above 100 people per square kilometer. ## Evaluation and Reported Performance - Precision was measured against the Groundsource dataset, but raw precision likely understates actual performance because some genuine floods are never reported. - A manual review of 100 alerts per continent found that many apparent false positives were confirmed flood events. - Recall was also evaluated against major floods recorded by the Global Disaster Awareness and Coordination System (GDACS). - Results indicate comparable precision and recall in regions such as South America and Southeast Asia and in wealthier countries with better instrumentation. The approach demonstrates how AI and unstructured public information can help provide scalable flash flood warnings where conventional monitoring infrastructure is limited. Its current urban focus and 20-kilometer resolution make it a broad early-warning tool rather than a replacement for highly localized sensor networks.

Read original(opens in new tab)

Machine LearningAIData AnalyticsFlood Forecasting+4

google·Mar 6, 2026·3 min readCurated summary

WAXAL: A large-scale open resource for African language speech technology

WAXAL is an open speech dataset designed to address the severe lack of technology resources for African languages. Covering 27 Sub-Saharan African languages spoken by more than 100 million people, it offers 1,846 hours of transcribed speech for ASR and over 565 hours of high-quality recordings for TTS under the permissive CC-BY-4.0 license. Built with African academic and community organizations, WAXAL aims to support inclusive speech systems while preserving and expanding access to the continent’s linguistic diversity. ## The Need for African-Language Speech Data - Voice assistants and transcription tools primarily support high-resource languages. - Hundreds of millions of African speakers remain underserved, despite Sub-Saharan Africa having more than 2,000 languages. - Limited training data makes it difficult to build accurate systems for tonal, morphologically complex, and code-switching languages. - WAXAL was developed beginning in 2021 to provide openly accessible resources for research and commercial innovation. ## WAXAL-ASR: Natural, Spontaneous Speech - Contains approximately 1,846 hours of transcribed audio. - Participants described image prompts across more than 50 topics rather than reading prepared scripts. - The method captured: - Natural linguistic variation - Tonal distinctions - Spontaneous speech patterns - Code-switching - Image-based elicitation produced more realistic speech data than conventional scripted recordings. ## WAXAL-TTS: High-Quality Voice Recordings - Includes more than 565 hours of phonetically balanced, high-fidelity audio. - Community members collaborated in pairs to create scripts of 10,000–20,000 words and alternate between reading and recording. - Some participants built custom studio boxes to improve recording quality. - Audio was segmented, aligned with transcripts, and reviewed for accuracy. - The clean recordings support natural-sounding synthetic voices and conversational systems. ## Supporting Full-Duplex Conversational Systems - The ASR data represents varied, real-world speech input. - The TTS data provides high-quality material for generating clear and natural speech. - Together, the datasets can support systems that both understand and respond using African languages. ## Community-Led Development - African universities and community organizations led the data collection, with guidance from Google Research. - Makerere University contributed data for nine languages. - The University of Ghana worked on eight languages, including image-prompted ASR collection. - Digital Umuganda and Addis Ababa University supported ASR collection for regional languages. - Media Trust, Loud n Clear, and the African Institute for Mathematical Sciences Senegal led studio-quality TTS recordings. - Partners retain ownership of the data while making it openly available under CC-BY-4.0. ## Research Enabled by the Project - A community-driven impaired-speech collection cookbook produced an open Akan dataset covering conditions such as cerebral palsy and stammering. - A separate project introduced a 5,000-hour corpus for five Ghanaian languages: Akan, Ewe, Dagbani, Dagaare, and Ikposo. - Benchmarking of Whisper, XLS-R, MMS, and W2v-BERT across 13 African languages showed that scaling benefits depend on linguistic complexity and domain alignment. - A literature review cataloged 74 datasets spanning 111 African languages and emphasized the need for conversational, multi-domain data and metrics such as Character Error Rate (CER). WAXAL provides a strong foundation for building more accurate and inclusive African-language speech technologies. Researchers and developers should use the openly licensed corpus, contribute derivative work, and help extend the collection to additional languages and communities.

Read original(opens in new tab)

Machine LearningData CollectionWaxalAsr+4

google·Feb 9, 2026·3 min readCurated summary

How AI trained on birds is surfacing underwater mysteries

Perch 2.0, Google DeepMind’s bioacoustics foundation model, was trained mainly on birds and terrestrial animals yet performs strongly on underwater audio. The study shows that its learned audio embeddings can support accurate whale, dolphin, reef-sound, and killer-whale classification with only a few labeled examples. This suggests that large, broadly trained bioacoustics models can transfer across environments and accelerate marine research without requiring extensive underwater training data. ## Underwater Mysteries and Bioacoustics - Ocean recordings reveal animal behavior, species distributions, and unexplained sounds. - The “biotwang,” recently attributed by NOAA to Bryde’s whales, illustrates how new calls and species identifications continue to emerge. - Google has previously developed models for humpback whales and multi-species whale detection. - Perch 2.0 extends this work despite having no underwater audio in its training data. ## Transfer Learning for Custom Classifiers - Researchers can use an existing model directly when its labels match their data. - For new sounds or datasets, transfer learning avoids training a deep neural network from scratch. - Perch 2.0 converts audio windows into compact numerical embeddings. - A logistic regression classifier is then trained on those embeddings using labeled examples. - This requires far less computation, experimentation, and training data than full neural-network training. ## Evaluation on Marine Datasets - The researchers tested Perch 2.0 with few-shot linear probes using 4, 8, 16, or 32 examples per class. - Performance was measured using ROC-AUC, where values closer to 1 indicate better class separation. - Evaluation datasets included: - **NOAA PIPAN:** Baleen-whale recordings, including minke, humpback, sei, blue, fin, and Bryde’s whales. - **ReefSet:** Reef biological sounds, fish, dolphins, anthropogenic noise, and waves. - **DCLDE:** Killer whales, humpbacks, abiotic sounds, unknown sounds, and killer-whale ecotypes. - More examples generally improved results. - ReefSet performance was already high with four examples per class for most models. - Perch 2.0 was consistently among the best-performing models across datasets and sample sizes. ## Comparisons with Other Models - Perch 2.0 was compared with Perch 1.0, SurfPerch, and Google’s multi-species whale model. - It also outperformed AVES-bird and AVES-bio on most underwater tasks. - The results show that strong underwater transfer is not limited to models trained on marine audio. ## Why Bird-Based Training Transfers to Whales - The authors suggest that large models trained on extensive datasets can generalize effectively to unfamiliar downstream tasks. - Shared acoustic patterns across animal vocalizations may allow representations learned from birds and other terrestrial species to remain useful underwater. - The findings challenge the assumption that a model must be trained directly on underwater recordings to perform well on marine classification tasks. ## Practical Tools for Researchers - Google provides a paper and a Google Colab tutorial. - The tutorial demonstrates an end-to-end workflow for building a whale-vocalization classifier. - It uses NOAA’s NCEI Passive Acoustic Data Archive and Google Cloud. - Researchers can create agile, task-specific models with relatively small labeled datasets. Perch 2.0 demonstrates that broad bioacoustic pretraining can substantially reduce the effort required to study marine sounds. Researchers can begin with general-purpose embeddings and adapt them to new whale species, calls, or underwater sound categories using only modest labeled data.

Read original(opens in new tab)

Machine LearningFoundation ModelsEmbeddingsFew Shot Learning+4

google·Jan 23, 2026·3 min readCurated summary

Introducing GIST: The Next Stage in Smart Sampling | Google Research

GIST is a data-subset selection algorithm designed to balance diversity and utility when training on massive datasets. It converts the difficult diversity–utility optimization problem into a series of thresholded graph problems and uses a bicriteria greedy strategy to find a high-quality subset efficiently. The algorithm guarantees at least half the value of the optimal solution, while the authors prove that improving beyond a 0.56 approximation is NP-hard. ## Why Smart Sampling Is Difficult - Large ML systems need to process datasets that are increasingly expensive to store, analyze, and train on. - Subset selection aims to choose a smaller but representative set of examples. - **Diversity** prevents redundant selections by maximizing the minimum distance between selected points, typically in embedding space. - **Utility** measures how much relevant or unique information the subset provides, modeled using monotone submodular functions. - Optimizing both objectives simultaneously is NP-hard: - A diversity-only method may select irrelevant examples. - A utility-only method may select many similar examples from one highly relevant cluster. ## How GIST Works ### Diversity Thresholding - GIST fixes a candidate minimum distance rather than optimizing the distance directly. - It builds a graph in which two data points are connected when they are closer than the chosen threshold. - Connected points are considered too similar to coexist in the selected subset. - Selecting points that are not connected enforces the desired spacing between examples. ### Utility-Constrained Independent Sets - For each threshold, GIST seeks a high-utility independent set: a group of points with no edges between them. - This corresponds to selecting valuable examples without choosing mutually conflicting or overly similar points. - Because maximum independent set is NP-complete and lacks practical general-purpose approximation algorithms, GIST uses a specially designed bicriteria greedy method. - The algorithm repeatedly selects high-scoring points and excludes nearby candidates, effectively creating “no-go zones” around selected data. ### Searching Across Thresholds - GIST evaluates all relevant distance thresholds derived from the dataset. - It greedily constructs a candidate subset for each threshold. - It returns the best candidate found across these runs. - If the optimal solution achieves minimum distance \(d\), GIST obtains comparable utility while guaranteeing a minimum distance of roughly \(d/2\). ## Theoretical Guarantees - GIST is presented as the first algorithm with a strong provable guarantee for this diversity–utility tradeoff. - Its output has at least half the value of the absolute optimum. - The authors also prove that finding a solution worth more than 0.56 of the optimum is NP-hard. - These results provide a mathematical guarantee that GIST is not merely producing empirically good subsets, but making a bounded tradeoff between informativeness and coverage. ## Practical Evaluation - GIST was evaluated against several common subset-selection approaches in ML applications. - Comparisons included: - **Random**, a simple baseline that often provides reasonable diversity. - **Margin**, which selects examples the model is uncertain about but does not explicitly promote diversity. - **k-center**, which minimizes representation gaps by keeping all data points close to a selected representative. - **Submod**, which combines utility with an older formulation of diversity. - The experiments, including image-classification benchmarks, reportedly show that GIST outperforms state-of-the-art alternatives while retaining formal guarantees. GIST is therefore a practical choice when subset selection must preserve both broad data coverage and task relevance. Its main advantage is combining competitive real-world performance with a clear approximation guarantee, rather than relying solely on heuristic results.

Read original(opens in new tab)

Machine LearningApproximation AlgorithmsSubset SelectionData Sampling+4