AI News from Google Research Blog

Latest coverage from Google Research Blog, summarized and scored for signal.

• Seeing beyond BMI: Estimating cardiometabolic risk with smartphone imagery — Google's PhotoScan deep learning model estimates body composition and predicts insulin resistance from standard smartphone photos, achieving accuracy comparable to expensive DXA clinical scans — potentially enabling metabolic risk screening without clinical infrastructure.
• Empty shelves or lost keys? Recall is the bottleneck for parametric factuality — Google Research's knowledge profiling framework finds frontier LLMs (Gemini 3, GPT-5) encode nearly all facts correctly but fail to recall many of them — most factual errors are retrieval failures, not knowledge gaps, shifting the optimal fix toward post-training and inference-time methods rather than more pretraining data.
• Advancing AMIE towards expert-level audio-visual clinical consultations — Google's AMIE medical AI now conducts real-time video clinical consultations, achieving expert-level performance in a first-of-its-kind randomized controlled study—interpreting visual cues like patient gait and breathing that text-based systems inherently discard.
• SymptomAI: Towards a conversational AI agent for everyday symptom assessment — Google Research's SymptomAI used Gemini Flash 2.0 to conduct end-to-end symptom interviews for 13,917 real patients in a randomized national study, with diagnostic accuracy validated against clinician outcomes two weeks later — the first AI symptom-checker evaluated in true real-world conditions rather than synthetic case studies.
• Towards a quantum computer that learns from its errors — Google Quantum AI published in Nature a reinforcement learning framework that lets a quantum computer continuously recalibrate its control parameters mid-computation — eliminating mandatory calibration halts that have capped quantum algorithm runtimes to minutes.
• Towards demystifying the creativity of diffusion models — Google Research's ICLR 2026 paper proves that diffusion models' ability to generate novel images is a mathematical consequence of neural networks learning a 'smoothed' score function — forcing interpolation between training examples along the data manifold rather than memorizing any single one.
• SensorFM: Towards a general intelligence and interface for wearable health data — Google Research released SensorFM, a foundation model pre-trained on over one trillion minutes of wearable sensor data from 5 million people — the largest wearable health dataset ever used — achieving state-of-the-art transfer to 35 health prediction tasks including cardiovascular, sleep, and mental health.
• The power of collaboration: How we can reduce traffic congestion — A six-month switchback experiment across 10 US cities shows that coordinating a small fraction of Google Maps users onto alternative routes reduces overall network congestion for all drivers — published in Nature Cities.
• Expanding our Heat Resilience data to 50+ global cities — Google Research expanded its AI-driven heat resilience dataset from 14 to 50+ global cities, mapping building-level rooftop reflectivity using Sentinel-2 fused with 30cm Airbus imagery — published in Nature Communications and accessible via a new Earth Engine App.
• Introducing TabFM: A zero-shot foundation model for tabular data — Google launches TabFM, a zero-shot foundation model for tabular classification and regression integrated directly into BigQuery ML, eliminating per-dataset model training and feature engineering with a single forward pass.
• Accelerating Gemini Nano models on Pixel with frozen Multi-Token Prediction — Google retrofitted Multi-Token Prediction onto frozen Gemini Nano v3 models for Pixel 9 and 10 using a late-exit strategy, generating multiple tokens per forward pass without a separate drafter model — speeding up AI Notification Summaries and Proofread while cutting energy use.
• Optimizing cloud economics with linear elastic caching — Google Research published a CIDR paper on linear elastic caching—a technique that frames cache eviction as a ski rental problem, using lightweight ML to dynamically resize cache allocation, targeting serverless environments where memory costs up to $3/day per GiB.
• Thinking to recall: How reasoning unlocks parametric knowledge in LLMs — Google Research finds chain-of-thought reasoning unlocks correct factual recall in LLMs even for simple single-hop questions — operating via two mechanisms: a computational buffer effect where intermediate tokens do latent work, and factual priming where related facts trigger the target answer.
• From pixels to planning: Earth AI for nature restoration — Google Research releases Vectorized Farmscapes 2020 — a dataset converting high-resolution satellite maps into actionable vector inventories of England's hedgerows, copses, and stone walls, enabling carbon accounting for fine-scale vegetation features that are too small for standard satellite detection.
• Research into how AI can help users understand skin conditions — Google Research published two papers in JAMA Dermatology showing that AI-assisted dermatology tools improve laypeople's ability to identify skin conditions and determine next steps — a study of 2,345 participants demonstrated measurable gains over searching without AI.
• A low-carbon computing platform from your retired phones — UC San Diego researchers, backed by Google, are deploying 2,000 retired Pixel phone motherboards as a research datacenter — showing smartphone cores match modern server single-thread performance and can cut embodied carbon without new hardware manufacturing.
• New framework for auditing machine unlearning — Google Research published Regularized f-Divergence Kernel Tests (AISTATS 2026), a new framework for auditing machine unlearning that maintains statistical power as models grow — addressing a critical gap in GDPR compliance verification tools.
• Unlocking dependable responses with Gemini Enterprise Agent Platform’s Agentic RAG — Google Research's agentic RAG framework for Gemini Enterprise Agent Platform uses multi-agent query planning and iterative cross-corpus retrieval to improve factuality by up to 34% over standard single-step RAG on complex enterprise queries.
• Towards passive heart health monitoring via smartphone camera — Google Research published a Nature paper on PHRM, a passive heart rate monitor that uses the front-facing smartphone camera during normal use — achieving wearable-level accuracy (< 5 bpm resting heart rate error) without any deliberate user action, and releasing the largest public facial video dataset for health research.
• The next chapter in flood resilience: Open sourcing Google’s hydrology framework — Google Research open-sourced its AI hydrology framework — the same architecture powering Google's Flood Hub riverine forecasts — enabling meteorological agencies and researchers to train localized flood forecasting models with their own data.
• A New Era of Innovation: Google Research at I/O 2026 — Google unveiled Gemini for Science at I/O 2026 — anchored by ERA and Co-Scientist, two systems published in Nature the same week — with demonstrated applications in hospital admission forecasting, California river basin runoff prediction, and antimicrobial resistance research.
• Private analytics via zero-trust aggregation — Google Research published a private analytics framework for monitoring on-device AI models that combines a new cryptographic aggregation protocol with trusted execution environments — allowing population-level insights without any single party, including Google, accessing individual user data.
• Empirical Research Assistance (ERA): From Nature publication to catalyzing Computational Discovery — Google's ERA (Empirical Research Assistance) AI tool achieves expert-level scientific coding performance across genomics, neuroscience, public health, and mathematics benchmarks in a Nature paper published today, and is now accessible via Google Labs' Computational Discovery.
• Catalyzing scientific impact through global partnerships and open resources — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Four ways Google Research scientists have been using Empirical Research Assistance — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.