Google Research describes results from Antigravity's /teamwork multiagent system across mathematics, chemistry, systems software, and ML engineering. The post explains design choices such as partitioned context, dedicated falsifiers like the Lean 4 checker, and verified outcomes including a Courtade–Kumar conjecture proof and a Rust port of libgav1. It also notes which tools are available to external researchers and developers.
Google Research engineering blog
Loading
Google Research presents an investigational framework that derives a modifiable coachable health age from passive wearable signals such as resting heart rate, steps, sleep, and Active Zone Minutes. A action-conditioned world model (SensorWM), trained on about 2.28 million person-weeks of Fitbit data, simulates 12-week habit scenarios, and the framework extends to blood biomarkers for organ-specific cardiovascular-kidney-metabolic profiling. Blinded evaluations with coaches and clinicians are...
Google Research and Google DeepMind introduce Language Explorer, an interactive web platform for visualizing global linguistic diversity, built on the open-source LinguaMeta dataset and part of the 1,000 Languages Initiative. The post also announces its installation in the Museum of Communication Berlin exhibition, with an online companion. It contains little technical detail about the tool's design or implementation.
Google Research describes AI-PPG-ACC-HTN, a multimodal model that uses consumer smartwatch PPG and accelerometer signals to passively detect signs of hypertension without cuff calibration. The system uses self-supervised contrastive pretraining on hundreds of thousands of hours of wearable data, aggregates embeddings over multiple days, and was validated against 24-hour ambulatory blood pressure monitoring in a 196-participant multi-center study. Reported sensitivity, specificity, and PPV are...
Google Research introduces Agile Deliberation, a framework for building AI image classifiers for subjective categories by treating classifier creation as an iterative conversation. It combines concept scoping into sub-categories and exclusions with VLM-driven mining of borderline images and an automatic prompt optimization loop, evaluated in 18 live 1.5-hour user sessions against zero-shot, modeling-collaborator, and manual baselines.
Google Research introduces a sleep framework for LLM continual learning, alternating active interaction with an offline sleep phase. During sleep, short-term context is consolidated into expanded MoE parameter blocks via generalized knowledge distillation and imitation learning, while a dreaming stage uses random expert routing and gradient-based sample filtering for self-improvement. The authors report gains over ICL and SFT baselines on translation, long-context (up to 10M tokens), factual...
Google Research describes how deep learning chest X-ray models for TB screening moved from 2019 benchmarks through regulatory certification with Nexus AI to deployment in about 40 facilities across six countries. The post covers open models (MedSigLIP, MedGemma) and HeAR cough-acoustic screening, and it explains the translation gap of offline operation, varied X-ray hardware, and medical device regulation.
A prospective multi-center study evaluates an AI model that estimates gestational age from blind sweep ultrasound scans performed by novice operators with low-cost probes. The model, fine-tuned on a small Chicago pilot dataset, achieved a mean absolute error of 4.18 days versus 4.48 days for expert sonographers and generalized to Nairobi without local tuning. The post also shares deployment lessons, including battery design and confidence-based rejection of poor-quality sweeps.
Google Research describes VERID, a semantic digital twin that lets AI agents plan, simulate, and safely execute liquid-handling protocols on lab robots. The post reports closed-loop aptamer design across three rounds, where agents diagnosed and fixed wet-lab failure modes such as G-quadruplex non-specific binding. It also covers an agent-guided epitope excision pipeline that accelerated structure prediction and yielded single-digit nanomolar FAP binders.
Google Research is launching Deep Review, an agentic AI system for verifying technical documents that originated as the Paper Assistant Tool used in pilots with STOC, ICML, NeurIPS, and ICLR. The post describes a four-stage pipeline of segmentation, adaptive compute budgeting, parallel section verification, and global synthesis with external grounding checks, and argues that single-prompt LLM calls fail on long documents. It also covers the public access tiers, API, and input-format limits.
Google Research describes a Geospatial Reasoning agent that pairs Gemini with Earth AI foundation models, modular spatial skills, and an upfront planning stage. On new GeoExpertQueries and GeoKTools benchmarks, the authors report that domain tools and planning outperform standalone LLMs, and they note that the autorater measures workflow quality rather than final-answer correctness.
Google Research's Research@ 2026 roundup presents ten projects spanning geospatial foundation models (Earth AI), a multi-agent orchestrator for reproducing scientific methods, the Deep Review agentic document-verification system, and a low-power Edge AI chip. It also covers TimesFM and TabFM for numerical forecasting, MedGemma and AMIE for health AI, Groundsource for disaster datasets, generative UI for education, and Play with Putty for collaborative prototyping. Most items are brief...
Researchers report results from a three-month randomized trial of AI-assisted patent drafting involving 133 lawyers. AI improved drafting quality for all participants, but only senior lawyers showed stronger unassisted judgment after 90 days, while junior outcomes became more variable.
Google Research demonstrates how the Population Dynamics Foundation Model converts privacy-preserving search, mobility, environmental, and built-environment signals into reusable location embeddings. Five global health case studies show gains in vaccination modeling, disease nowcasting, dengue forecasting, postpartum-depression risk prediction, and cholera outbreak warnings.
Google Research examines privacy and security challenges in autonomous AI agents through the lens of Contextual Integrity. The report proposes contextual policy engines, layered system and model safeguards, user controls, multi-agent guardrails, and dynamic benchmarks for evaluating agent behavior.
Google Research presents a TEE-based federated learning system that provides externally verifiable privacy guarantees through encrypted uploads, access policies, remote attestation, differential privacy, and reproducible builds. The architecture shifts training computation to servers, improving device coverage, training speed, and model accuracy while supporting fault-tolerant recovery.
Google Research introduces Diffusion Controller, a lightweight control network that steers image-generation models toward better prompt and preference alignment while keeping the base model frozen. The framework unifies inference-time guidance and fine-tuning, supports restricted-access models, and uses PPO or reward-weighted loss to improve quality without destabilizing generation.
Google researchers present a unified multi-agent framework for generating coherent, minutes-long video narratives. Its components use hierarchical planning, persistent visual memory, adaptive interpolation and extrapolation, and closed-loop prompt refinement to reduce character drift, scene inconsistency, and cascading pipeline failures.
Google Research introduces MilleMiglia, an open-source C++ generator for realistic, privacy-preserving middle-mile logistics benchmarks. It models space-time flows, fixed vehicle schedules, distribution-center capacity, shipment synchronization, and scalable instances for optimization and machine-learning research.
Google Research presents a generative UI system that helps teachers create curriculum-aligned, interactive STEM simulations. The approach combines structured learning objectives, progressively difficult game levels, AI-generated scaffolding, iterative agentic evaluation, and teacher validation to improve educational quality and solvability.