Google researchers introduce a method to retrofit Multi-Token Prediction onto deployed Gemini Nano v3 models without retraining the backbone, achieving faster inference and lower energy consumption on mobile devices. Deployed on Pixel 9 and 10 series, it boosts speed by over 50% for features like AI Notification Summaries and Proofread.
Freezes the backbone and attaches a lightweight MTP head, enabling seamless acceleration without the memory overhead of a separate drafter model.
Zero-copy architecture allows the MTP head to leverage the main model's KV cache directly, reducing memory usage by 130MB and eliminating draft prefill latency.
Google researchers propose linear elastic caching, which models cache management as a ski rental problem, using lightweight machine learning to dynamically adjust cache size. In Spanner production, it reduced memory usage by 15.5%, TCO by ~5%, with only 5.5% more cache misses and negligible I/O impact.
Linear elastic caching treats memory cost as a continuous variable, dynamically adjusting cache size instead of fixed allocation.
Based on the ski rental problem, it decides for each data item whether to 'rent' memory or 'buy' a cache miss.
Google Research reveals a counterintuitive phenomenon: even for simple factual questions, prompting LLMs to generate reasoning chains improves answer accuracy. Two mechanisms are identified: computational buffer (extra tokens provide additional computation) and factual priming (generating related facts facilitates retrieval).
Reasoning helps models recall simple facts that are otherwise unreachable, even without step-by-step reasoning.
Mechanism 1: Computational buffer — generating meaningless reasoning tokens also provides extra computation, improving recall.
Google Research presents two studies on dermatology AI tools. A large-scale survey found AI assistance tripled users' accuracy in naming skin conditions, but improving next-step decisions remains challenging. A qualitative community study showed the app helped users and clinicians, with 92% of clinicians finding it helpful.
AI assistance nearly tripled the accuracy of users in naming skin conditions (from 8% to 23%).
Determining appropriate next steps (e.g., home remedy vs. clinic visit) showed no significant improvement with AI.
Google researchers introduce Regularized f-Divergence Kernel Tests to audit machine unlearning and privacy. The framework adaptively selects optimal divergence measures, improving detection of data leaks and unlearning failures while requiring fewer samples and less tuning.
Two-sample tests lose power for large models; new framework is more sensitive and adaptive.
Uses f-divergences (chi-squared, KL, hockey-stick) to detect both global and local data shifts.
Google's new Agentic RAG framework uses multiple specialized agents to iteratively search and verify context before answering complex queries, achieving up to 34% higher accuracy than standard RAG.
Multi-agent architecture with Planner, Query Rewriter, and Sufficient Context Agent
Iterative retrieval until context is complete, reducing guesswork
Researchers at Google have developed a system called PHRM that passively measures heart rate and resting heart rate using the front-facing camera of a smartphone during everyday use. In a study published in Nature, the system achieved an accuracy of less than 10% mean absolute percentage error compared to ECG, and less than 5 bpm error for daily resting heart rate compared to a wearable. The system was tested on a diverse dataset of over 350,000 video clips from nearly 700 participants, ensuring balanced representation across skin tones. PHRM outperformed 15 leading remote photoplethysmography models and is the only model to meet accuracy standards for all skin tones in real-world conditions.
Google's PHRM system uses the smartphone's front-facing camera to passively monitor heart rate and resting heart rate after face unlock events.
In a Nature study, PHRM achieved <10% MAPE for heart rate vs ECG and <5 bpm MAE for daily resting heart rate vs a wearable, across all skin tones.
At Google I/O 2026, Google Research showcased breakthroughs in scientific discovery, health, edge computing, and weather prediction. Highlights include Gemini for Science (ERA, Co-Scientist), Google Health app, Symptom AI, AMIE, Coral NPU, and AI for extreme weather. These innovations demonstrate AI's potential to amplify human ingenuity.
Google launched Gemini for Science with ERA and Co-Scientist to accelerate scientific discovery.
Health advancements include Google Health app, Symptom AI, and AMIE improving clinical care.
Published today in Nature, ERA is an AI tool that uses Gemini to write and optimize scientific code, achieving expert-level performance across benchmarks. It helped build Computational Discovery, now available through a trusted tester program in Google Labs. Applications include epidemiology, hydrology, CO2 mapping, solar energy, and retail forecasting.
ERA uses Gemini to write and optimize scientific code, addressing the iterative testing bottleneck in research.
Nature publication shows expert-level results across genomics, public health, neuroscience, and more.
Google Research's open science approach leverages open-source software and datasets, partnering with global institutions to drive breakthroughs in genomics, neuroscience, climate, biodiversity, and healthcare. The article details tools like DeepVariant, Neuroglancer, Open Buildings, SpeciesNet, and HAI-DEF, with real-world examples of impact on farmers, patients, and conservation efforts.
Google Research collaborates with numerous global organizations to advance open science.
Open tools and datasets empower over 250,000 researchers worldwide across multiple scientific domains.
Since introducing Empirical Research Assistance in fall 2025, Google Research scientists have applied it to real-world problems in epidemiology, cosmology, atmospheric monitoring, and neuroscience, demonstrating AI's transformative potential to accelerate scientific discovery.
ERA matches or outperforms CDC tools in forecasting flu, COVID-19, and RSV hospitalizations.
Combined with Gemini Deep Think, ERA derived solutions for cosmic string gravitational radiation.
Google Photos' Auto frame feature uses ML to interpret photos as 3D scenes and automatically adjust the camera angle to generate a new perspective. The two-stage method involves 3D point map estimation and generative inpainting, enhancing portraits by fixing perspective distortion.
Google Photos introduces Auto frame to automatically recompose photos using AI.
The technology treats 2D photos as 3D scenes, adjusting camera angles for novel views.
Google Cloud researchers introduce ReasoningBank, a novel agent memory framework that distills generalizable reasoning strategies from both successful and failed experiences, enabling agents to continuously learn after deployment. It outperforms baselines on web and software engineering benchmarks.
ReasoningBank structures memories into Title, Description, and Content from past experiences.
Memory-aware test-time scaling (MaTTS) leverages parallel and sequential scaling to further boost performance.
Google introduces Simula, a reasoning-first framework that treats synthetic data generation as mechanism design, enabling fine-grained control over coverage, complexity, and quality for specialized AI domains.
Simula reframes synthetic data generation as dataset-level mechanism design.
It decomposes generation into four controllable axes: global and local diversification, complexification, and quality checks.
Google Research developed MoGen, a model that generates synthetic neuron morphologies using AI, improving brain reconstruction accuracy by 4.4%. For a full mouse brain, this translates to saving 157 person-years of manual proofreading.
MoGen generates realistic 3D neuron shapes via point cloud flow matching, enhancing training data.
Adding 10% synthetic data to the PATHFINDER model reduces reconstruction errors by 4.4%, mainly by reducing merge errors.
Google Research and New York University have developed Vantage, a system that uses generative AI to assess future-ready skills like collaboration and critical thinking. It simulates conversations with AI avatars steered by an Executive LLM, and evaluation shows AI scoring matches human experts. Vantage is now available on Google Labs.
Vantage uses generative AI to create simulated environments for assessing future-ready skills.
An Executive LLM steers AI avatars to introduce challenges and gather evidence of skills.
Google Research unveils ConvApparel, a dataset and evaluation framework to quantify the realism gap in LLM-based user simulators. With dual-agent data collection and three-pillar validation, it shows data-driven simulators outperform prompt-based ones but a gap remains.
ConvApparel includes over 4,000 human-AI conversations with both helpful and unhelpful agents.
Evaluation uses population alignment, human-likeness scoring, and counterfactual validation.
Google Cloud researchers introduce PaperVizAgent for generating publication-ready figures and ScholarPeer for automated, rigorous peer review. Both systems outperform existing baselines by significant margins, demonstrating the potential of multi-agent AI in academic research.
PaperVizAgent generates high-quality academic figures from text using a five-agent collaborative system with iterative refinement.
ScholarPeer emulates senior reviewers through context acquisition, active verification, and multi-aspect questioning.
Google Research introduces a systematic evaluation framework that transforms established psychological questionnaires into situational judgment tests for LLMs. Testing 25 models reveals gaps in alignment: models deviate from human consensus in high-agreement scenarios and are overconfident when opinions are divided.
New framework adapts psychological questionnaires into situational judgment tests to assess LLM behavioral dispositions in realistic scenarios.
25 LLMs show near-perfect alignment when human consensus is unanimous, but plateau at 80% when consensus is below 90%.