google·Sep 10, 2025Original article

Speculative cascades — A hybrid approach for smarter, faster LLM inference (opens in new tab)

Speculative cascades represent a hybrid inference method that integrates the cost-efficiency of model cascades with the latency-reducing benefits of speculative decoding. By utilizing a smaller drafter model to generate token sequences that are verified in parallel by a larger expert model, this approach allows for high-speed generation while maintaining flexible quality standards. The result is a system that achieves superior cost-quality trade-offs and higher speed-ups than either traditional cascading or standard speculative decoding alone. ### Limitations of Cascades and Speculative Decoding * **Sequential Bottlenecks in Cascades:** Traditional cascades use a deferral rule to decide if a small model can handle a prompt. If the small model is not confident, the system waits for it to finish before starting the large model from scratch, wasting significant time. * **Strict Matching in Speculative Decoding:** This method requires the large model to verify the small model’s tokens. Even if the small model produces a factually correct and high-quality response, the large model will reject the entire draft if the tokens do not match its own preferred output exactly. * **Trade-off Divergence:** Cascades prioritize reducing computational costs but suffer from latency when deferring, while speculative decoding prioritizes speed but often performs redundant work because it mandates identical output to the larger model. ### The Speculative Cascades Mechanism * **Parallel Verification with Deferral:** Speculative cascades use the parallel processing of speculative decoding but introduce a flexible decision rule. The system can choose to accept the smaller model’s draft even if it differs from the larger model’s prediction, provided it meets a confidence threshold. * **Flexible Token Matching:** Unlike standard speculative decoding, which often relies on strict token-by-token matching, speculative cascades allow for "probabilistic matches" or quality-based acceptance to prevent unnecessary rejections. * **Resource Optimization:** By strategically deferring to the smaller model for certain segments of the generation, the system reduces the total work required from the expensive expert model without losing the speed of parallel execution. ### Empirical Results and Performance * **Model Testing:** The approach was validated using Gemma and T5 models across diverse language tasks, including reasoning, coding, translation, and question answering. * **Superior Trade-offs:** Testing showed that speculative cascades consistently outperformed baselines in cost-quality metrics, providing faster inference without the strict "all-or-nothing" quality constraints of speculative decoding. * **Task Versatility:** The hybrid method proved effective across both creative tasks (like summarization) and factual tasks (like math or coding), where different levels of "correctness" are acceptable. Speculative cascades offer a practical path for scaling LLM deployments by balancing the high cost of large models with the need for low-latency user experiences. Developers looking to optimize inference should consider this hybrid approach to capture the efficiency of small models while retaining the oversight of larger, more capable ones.

AILarge Language ModelsGemmaInference Optimization+5

google·Sep 8, 2025Original article

Accelerating scientific discovery with AI-powered empirical software (opens in new tab)

Google Research has introduced an AI-powered system designed to accelerate scientific discovery by automating the creation and optimization of "empirical software." By leveraging the Gemini model and tree search optimization, the system can propose, implement, and iteratively improve code for complex multidisciplinary challenges, achieving results that match or exceed human expert performance. This approach transforms scientific hypothesis evaluation from a months-long manual coding process into an automated search that can be completed in hours or days. ### The Concept of Empirical Software and Scorable Tasks * The system shifts focus from traditional functional correctness to "empirical software," where the primary objective is to maximize a predefined quality score. * It targets "scorable tasks," which are defined by a problem description, a specific scoring metric, and a dataset for training and validation. * This framework addresses the research bottleneck where scientists must manually test hundreds of models or parameters to achieve a breakthrough. ### System Architecture and Optimization Strategy * The engine takes a task description and optional context—such as ideas from scientific literature—as input to generate novel methodological concepts. * It utilizes a tree search strategy inspired by AlphaZero, employing an upper confidence bound to navigate and prioritize thousands of potential code variants. * The LLM acts as an iterative rewriter, refining executable code within a sandbox to continuously improve the performance score. * Outputs are designed to be fully verifiable, interpretable, and reproducible, providing scientists with the specific coded solutions used to reach a result. ### Demonstrated Performance Across Scientific Domains * The system was tested on six diverse benchmarks, including genomics, public health, geospatial analysis, neuroscience, and time-series forecasting. * In genomics, the system tackled the "batch integration" of single-cell RNA sequencing (scRNA-seq) data, a complex problem involving the removal of noise while preserving biological signals. * The AI discovered 40 novel methods that outperformed top expert-developed tools within the OpenProblems V2.0.0 batch integration benchmark. * Evaluation focused on advanced capabilities such as zero-shot generalization, high-dimensional signal processing, and uncertainty quantification. This system represents a significant shift toward "research engines" that participate actively in the scientific method through iterative experimentation. Scientists can utilize these tools to explore a much broader range of hypotheses than manual coding allows, potentially leading to faster breakthroughs in data-heavy fields like genomics and climate modeling.

AILarge Language ModelsGeminiCode Generation+4

google·Aug 26, 2025Original article

How Google’s AI can help transform health professions education (opens in new tab)

To address a projected global deficit of 11 million healthcare workers by 2030, Google Research is exploring how generative AI can provide personalized, competency-based education for medical professionals. By combining qualitative user-centered design with quantitative benchmarking of the pedagogically fine-tuned LearnLM model, researchers have demonstrated that AI can effectively mimic the behaviors of high-quality human tutors. The studies conclude that specialized models, now integrated into Gemini 2.5 Pro, can significantly enhance clinical reasoning and adapt to the individual learning styles of medical students. ## Learner-Centered Design and Participatory Research * Researchers conducted interdisciplinary co-design workshops featuring medical students, clinicians, and AI researchers to identify specific educational needs. * The team developed a rapid prototype of an AI tutor designed to guide learners through clinical reasoning exercises anchored in synthetic clinical vignettes. * Qualitative feedback from medical residents and students highlighted a demand for "preceptor-like" behaviors, such as the ability to manage cognitive load, provide constructive feedback, and encourage active reflection. * Analysis revealed that learners specifically value AI tools that can identify and bridge individual knowledge gaps rather than providing generic information. ## Quantitative Benchmarking via LearnLM * The study utilized LearnLM, a version of Gemini fine-tuned specifically for educational pedagogy, and compared its performance against Gemini 1.5 Pro. * Evaluations were conducted using 50 synthetic scenarios covering a spectrum of medical education, ranging from preclinical topics like platelet activation to clinical subjects such as neonatal jaundice. * Medical students engaged in 290 role-playing conversations, which were then evaluated based on four primary metrics: overall experience, meeting learning needs, enjoyability, and understandability. * Physician educators performed blinded reviews of conversation transcripts to assess whether the AI adhered to medical education standards and core competencies. ## Pedagogical Performance and Expert Evaluation * LearnLM was consistently rated higher than the base model by both students and educators, with experts noting it behaved "more like a very good human tutor." * The fine-tuned model demonstrated a superior ability to maintain a conversation plan and use grounding materials to provide accurate, context-aware instruction. * Findings suggest that pedagogical fine-tuning is essential for AI to move beyond simple fact-delivery and toward true interactive tutoring. * These specialized learning capabilities have been transitioned from the research phase into Gemini 2.5 Pro to support broader educational applications. By integrating these specialized AI behaviors into medical training pipelines, institutions can provide scalable, individualized support to students. The transition of LearnLM’s pedagogical features into Gemini 2.5 Pro provides a practical framework for developers to create tools that not only provide medical information but actively foster the critical thinking skills required for clinical practice.

AIGenerative AIGeminiFine Tuning+5

google·Aug 25, 2025Original article

A scalable framework for evaluating health language models (opens in new tab)

Researchers at Google have developed a scalable framework for evaluating health-focused language models by replacing subjective, high-complexity rubrics with granular, binary criteria. This "Adaptive Precise Boolean" approach addresses the high costs and low inter-rater reliability typically associated with expert-led evaluation in specialized medical domains. By dynamically filtering rubric questions based on context, the framework significantly improves both the speed and precision of model assessments. ## Limitations of Traditional Evaluation * Current evaluation practices for health LLMs rely heavily on human experts, making them cost-prohibitive and difficult to scale. * Standard tools, such as Likert scales (e.g., 1-5 ratings) or open-ended text, often lead to subjective interpretations and low inter-rater consistency. * Evaluating complex, personalized health data requires a level of detail that traditional broad-scale rubrics fail to capture accurately. ## Precise Boolean Rubrics * The framework "granularizes" complex evaluation targets into a larger set of focused, binary (Yes/No) questions. * This format reduces ambiguity by forcing raters to make definitive judgments on specific aspects of a model's response. * By removing the middle ground found in multi-point scales, the framework produces a more robust and actionable signal for programmatic model refinement. ## The Adaptive Filtering Mechanism * To prevent the high volume of binary questions from overwhelming human raters, the researchers introduced an "Adaptive" layer. * The framework uses the Gemini model as a zero-shot classifier to analyze the user query and LLM response, identifying only the most relevant rubric questions. * This data-driven adaptation ensures that human experts only spend time on pertinent criteria, resulting in "Human-Adaptive Precise Boolean" rubrics. ## Performance and Reliability Gains * The methodology was validated in the domain of metabolic health, covering topics like diabetes, obesity, and cardiovascular disease. * The Adaptive Precise Boolean approach reduced human evaluation time by over 50% compared to traditional Likert-scale methods. * Inter-rater reliability, measured through intra-class correlation coefficients (ICC), was significantly higher than the baseline, proving that simpler scoring can provide a higher quality signal. This framework demonstrates that breaking down complex medical evaluations into simple, machine-filtered binary questions is a more efficient path toward safe and accurate health AI. Organizations developing domain-specific models should consider adopting adaptive binary rubrics to balance the need for expert oversight with the requirements of large-scale model iteration.

AILarge Language ModelsGeminiNatural Language Processing+4

google·Aug 20, 2025Original article

From massive models to mobile magic: The tech behind YouTube real-time generative AI effects (opens in new tab)

YouTube has successfully deployed over 20 real-time generative AI effects by distilling the capabilities of massive cloud-based models into compact, mobile-ready architectures. By utilizing a "teacher-student" training paradigm, the system overcomes the computational bottlenecks of high-fidelity generative AI while ensuring the output remains responsive on mobile hardware. This approach allows for complex transformations, such as cartoon style transfer and makeup application, to run frame-by-frame on-device without sacrificing the user’s identity. ### Data Curation and Diversity * The foundation of the effects pipeline relies on high-quality, properly licensed face datasets. * Datasets are meticulously filtered to ensure a uniform distribution across different ages, genders, and skin tones. * The Monk Skin Tone Scale is used as a benchmark to ensure the effects work equitably for all users. ### The Teacher-Student Framework * **The Teacher:** A large, powerful pre-trained model (initially StyleGAN2 with StyleCLIP, later transitioning to Google DeepMind’s Imagen) acts as the "expert" that generates high-fidelity visual effects. * **The Student:** A lightweight UNet-based architecture designed for mobile efficiency. It utilizes a MobileNet backbone for both the encoder and decoder to ensure fast frame-by-frame processing. * The distillation process narrows the scope of the massive teacher model into a student model focused on a single, specific task. ### Iterative Distillation and Training * **Data Generation:** The teacher model processes thousands of images to create "before and after" pairs. These are augmented with synthetic elements like AR glasses, sunglasses, and hand occlusions to improve real-world robustness. * **Optimization:** The student model is trained using a sophisticated combination of loss functions, including L1, LPIPS, Adaptive, and Adversarial loss, to balance numerical accuracy with aesthetic quality. * **Architecture Search:** Neural architecture search is employed to tune "depth" and "width" multipliers, identifying the most efficient model structure for different mobile hardware constraints. ### Addressing the Inversion Problem * A major challenge in real-time effects is the "inversion problem," where the model struggles to represent a real face in latent space, leading to a loss of the user's identity (e.g., changes in skin tone or clothing). * YouTube uses Pivotal Tuning Inversion (PTI) to ensure that the user's specific features are preserved during the generative process. * By editing images in the latent space—a compressed numerical representation—the system can apply stylistic changes while maintaining the core characteristics of the original video stream. By combining advanced model distillation with on-device optimization via MediaPipe, YouTube demonstrates a practical path for bringing heavy generative AI research into consumer-facing mobile applications.

AIGenerative AIComputer VisionKnowledge Distillation+5

google·Aug 19, 2025Original article

Securing private data at scale with differentially private partition selection (opens in new tab)

Google Research has introduced a novel parallel algorithm called MaxAdaptiveDegree (MAD) to enhance differentially private (DP) partition selection, a critical process for identifying common data items in massive datasets without compromising individual privacy. By utilizing an adaptive weighting mechanism, the algorithm optimizes the utility-privacy trade-off, allowing researchers to safely release significantly more data than previous non-adaptive methods. This breakthrough enables privacy-preserving analysis on datasets containing hundreds of billions of items, scaling up to three orders of magnitude larger than existing sequential approaches. ## The Role of DP Partition Selection * DP partition selection identifies a meaningful subset of unique items from large collections based on their frequency across multiple users. * The process ensures that no single individual's data can be identified in the final list by adding controlled noise and filtering out items that are not sufficiently common. * This technique is a foundational step for various machine learning tasks, including extracting n-gram vocabularies for language models, analyzing private data streams, and increasing efficiency in private model fine-tuning. ## The Weight, Noise, and Filter Paradigm * The standard approach to private partition selection begins by computing a "weight" for each item, typically representing its frequency, while ensuring "low sensitivity" so no single user has an outsized impact. * Random Gaussian noise is added to these weights to obfuscate exact counts, preventing attackers from inferring the presence of specific individuals. * A threshold determined by DP parameters is then applied; only items whose noisy weights exceed this threshold are included in the final output. ## Improving Utility via Adaptive Weighting * Traditional non-adaptive methods often result in "wastage," where highly popular items receive significantly more weight than necessary to cross the selection threshold. * The MaxAdaptiveDegree (MAD) algorithm introduces adaptivity by identifying items with excess weight and rerouting that weight to "under-allocated" items sitting just below the threshold. * This strategic reallocation allows a larger number of less-frequent items to be safely released, significantly increasing the utility of the dataset without compromising privacy or computational efficiency. ## Scalability and Parallelization * Unlike sequential algorithms that process data one piece at a time, MAD is designed as a parallel algorithm to handle the scale of modern user-based datasets. * The algorithm can process datasets with hundreds of billions of items by breaking the problem down into smaller parts computed simultaneously across multiple processors. * Google has open-sourced the implementation on GitHub to provide the research community with a tool that maintains robust privacy guarantees even at a massive scale. Researchers and data scientists working with large-scale sensitive datasets should consider implementing the MaxAdaptiveDegree algorithm to maximize the amount of shareable data while strictly adhering to user-level differential privacy standards.

AIMachine LearningDifferential PrivacyData Privacy+4

google·Aug 13, 2025Original article

Beyond billion-parameter burdens: Unlocking data synthesis with a conditional generator (opens in new tab)

The CTCL (Data Synthesis with ConTrollability and CLustering) framework provides a lightweight alternative to the computationally expensive process of fine-tuning billion-parameter models for differentially private synthetic data generation. By utilizing a 140-million parameter generator and a universal topic model, the system achieves high-quality distribution matching while remaining accessible for resource-constrained applications. This approach allows for the generation of unlimited synthetic samples without incurring additional privacy costs, consistently outperforming existing API-based and large-scale baselines under strict privacy guarantees. ### Pre-training Universal Components The framework relies on two core components developed using large-scale public corpora, which can be reused across different private domains: * **CTCL-Topic:** A universal topic model derived from Wikipedia documents. It uses BERTopic to embed and cluster data into approximately 1,000 distinct topics, each represented by 10 descriptive keywords. * **CTCL-Generator:** A conditional language model based on the 140M-parameter BART-base architecture. It was pre-trained on 430 million description–document pairs from the SlimPajama dataset, with descriptions generated by Gemma-2-2B to ensure the model can generate text based on specific input conditions. ### Learning the Private Domain Once the universal components are established, the framework learns the specific characteristics of a private dataset through a two-step process: * **Differentially Private (DP) Histograms:** The system captures high-level distributional information by creating a DP-protected histogram that represents the percentage of each topic present in the private corpus. * **DP Fine-Tuning:** Each document in the private dataset is associated with its corresponding keywords from the CTCL-Topic model. The CTCL-Generator is then fine-tuned on these keyword-document pairs using differential privacy to ensure individual data points are protected. ### Controllable Data Generation The final stage involves producing the synthetic dataset by sampling from the fine-tuned generator: * **Proportional Sampling:** The system generates data by targeting the exact topic proportions found in the private domain histogram. * **Keyword Conditioning:** For each topic, the model uses the associated 10 keywords as input to prompt the DP fine-tuned generator to produce relevant documents. * **Post-Processing Efficiency:** Because the generator is already fine-tuned with DP, the framework can generate an unlimited number of synthetic samples without further privacy budget expenditure, a significant advantage over iterative selection algorithms. CTCL offers a highly scalable and efficient solution for organizations needing to synthesize private text data without the infrastructure requirements of massive LLMs. Its ability to maintain topic-wise distribution through keyword conditioning makes it an ideal choice for specialized domains where maintaining the statistical utility of the data is as critical as protecting user privacy.

AILarge Language ModelsFine TuningDifferential Privacy+5

google·Aug 11, 2025Original article

Enabling physician-centered oversight for AMIE (opens in new tab)

Guardrailed-AMIE (g-AMIE) is a diagnostic AI framework designed to perform patient history-taking while strictly adhering to safety guardrails that prevent it from providing direct medical advice. By decoupling data collection from clinical decision-making, the system enables an asynchronous oversight model where primary care physicians (PCPs) review and finalize AI-generated medical summaries. In virtual clinical trials, g-AMIE’s diagnostic outputs and patient communications were preferred by overseeing physicians and patient actors over human-led control groups. ## Multi-Agent Architecture and Guardrails * The system utilizes a multi-agent setup powered by Gemini 2.0 Flash, consisting of a dialogue agent, a guardrail agent, and a SOAP note agent. * The dialogue agent conducts history-taking in three distinct phases: general information gathering, targeted validation of a differential diagnosis, and a conclusion phase for patient questions. * A dedicated guardrail agent monitors and rephrases responses in real-time to ensure the AI abstains from sharing individualized diagnoses or treatment plans directly with the patient. * The SOAP note agent employs sequential multi-step generation to separate summarization tasks (Subjective and Objective) from more complex inferential tasks (Assessment and Plan). ## The Clinician Cockpit and Asynchronous Oversight * To facilitate human review, researchers developed the "clinician cockpit," a web interface co-designed with outpatient physicians through semi-structured interviews. * The interface is structured around the standard SOAP note format, presenting the patient’s perspective, measurable data, differential diagnosis, and proposed management strategy. * This framework allows overseeing PCPs to review cases asynchronously, editing the AI’s proposed differential diagnoses and management plans before sharing a final message with the patient. * The separation of history-taking from decision-making ensures that licensed medical professionals retain ultimate accountability for patient care. ## Performance Evaluation via Virtual OSCE * The system was evaluated in a randomized, blinded virtual Objective Structured Clinical Examination (OSCE) involving 60 case scenarios. * g-AMIE’s performance was compared against primary care physicians, nurse practitioners, and physician assistants who were required to operate under the same restrictive guardrails. * Overseeing PCPs and independent physician raters preferred g-AMIE’s diagnostic accuracy and management plans over those of the human control groups. * Patient actors reported a preference for the messages generated by g-AMIE compared to those drafted by human clinicians in the study. While g-AMIE demonstrates high potential for human-AI collaboration in diagnostics, the researchers emphasize that results should be interpreted with caution. The workflow was specifically optimized for AI characteristics, and human clinicians may require specialized training to perform effectively within such highly regulated guardrail frameworks.

AIGenerative AINatural Language ProcessingMulti Agent Systems+4

google·Aug 6, 2025Original article

Achieving 10,000x training data reduction with high-fidelity labels (opens in new tab)

Google Ads researchers have developed a scalable active learning curation process that reduces the volume of training data required for fine-tuning LLMs by up to four orders of magnitude. By iteratively identifying the most informative and diverse examples through clustering and expert review, the method achieves significantly higher human-model alignment than traditional large-scale crowdsourced datasets. This approach effectively addresses the high costs and complexities of classifying ambiguous content, such as unsafe ads, where high-fidelity data is scarce and concept drift is frequent. ### The Iterative Curation Process * **Initial Labeling:** The process begins with a zero- or few-shot model (LLM-0) that generates a large, typically imbalanced dataset of "positive" and "benign" labels. * **Clustering and Confusion Identification:** Separate clusters are created for each label set; overlapping clusters indicate areas where the model is confused. * **Expert Sampling:** Human experts review pairs of examples located near the decision boundary of these overlapping clusters, prioritizing those that cover a larger area of the search space to ensure diversity. * **Recursive Refinement:** Expert labels are split into fine-tuning and evaluation sets; the model is retrained and the process repeats until model-human alignment plateaus or matches internal expert agreement. ### Measuring Alignment via Cohen’s Kappa * **Metric Selection:** Because ad safety is often subjective, the researchers use Cohen’s Kappa instead of precision and recall to measure how well two independent annotators align beyond chance. * **Performance Benchmarks:** A Kappa value above 0.8 is considered exceptional, while 0.4 is the minimum for acceptability. * **Goal Alignment:** The curation process aims to move model performance toward the "ceiling" of internal human agreement (which measured between 0.78 and 0.81 in these experiments). ### Experimental Results and Efficiency * **Model Scaling:** Experiments involved fine-tuning Gemini Nano-1 (1.8B parameters) and Nano-2 (3.25B parameters) on tasks of varying complexity. * **Drastic Data Reduction:** The curated method reached performance plateaus using fewer than 500 expert-labeled examples, compared to a baseline of 100,000 crowdsourced labels. * **Quality Gains:** Despite using 10,000x less data, the curated models saw up to a 65% improvement in alignment with human experts over the crowdsourced baselines. * **Class Balancing:** The process naturally corrected for production imbalances, moving from <1% positive examples in raw traffic to ~40% in the final curated sets. This curation method is a highly effective strategy for organizations managing high-stakes classification tasks where "ground truth" is subjective or data curation is prohibitively expensive. By shifting focus from data quantity to the quality and diversity of examples at the decision boundary, developers can maintain high-performing models that adapt quickly to evolving safety policies.

AILarge Language ModelsMachine LearningGemini+4

google·Aug 5, 2025Original article

Insulin resistance prediction from wearables and routine blood biomarkers (opens in new tab)

Researchers at Google have developed a novel machine learning approach to predict insulin resistance (IR) by integrating wearable device data with routine blood biomarkers. This method aims to provide a scalable, less invasive alternative to traditional "gold standard" tests like the euglycemic insulin clamp or specialized HOMA-IR assessments. The study demonstrates that combining digital biomarkers with common laboratory results can effectively identify individuals at risk for type 2 diabetes, particularly within high-risk populations. ## Barriers to Early Diabetes Screening * Insulin resistance is a primary precursor to approximately 70% of type 2 diabetes cases, yet it often remains undetected until the disease has progressed. * Current diagnostic standards are frequently omitted from routine check-ups due to high costs, invasiveness, and the requirement for specific insulin blood tests that are not standard practice. * Early detection is vital because insulin resistance is often reversible through lifestyle modifications, making accessible screening tools a high priority for preventative medicine. ## The WEAR-ME Multimodal Dataset * The research utilized the "WEAR-ME" study, which collected data from 1,165 remote participants across the U.S. via the Google Health Studies app. * Digital biomarkers were gathered from Fitbit and Google Pixel Watch devices, tracking metrics such as resting heart rate, step counts, and sleep patterns. * Clinical data was provided through a partnership with Quest Diagnostics, focusing on routine blood biomarkers like fasting glucose and lipid panels, supplemented by participant surveys on diet, fitness, and demographics. ## Predictive Modeling and Performance * Deep neural network models were trained to estimate HOMA-IR scores by analyzing different combinations of the collected data streams. * While models using only wearables and demographics achieved an area under the receiver operating characteristic curve (auROC) of 0.70, adding fasting glucose data boosted the auROC to 0.78. * The most comprehensive models, which combined wearables, demographics, and full routine blood panels, achieved the highest accuracy across the study population. * Performance was notably strong in high-risk sub-groups, specifically individuals with obesity or sedentary lifestyles. ## AI-Driven Interpretation and Literacy * To assist with data translation, the researchers developed a prototype "Insulin Resistance Literacy and Understanding Agent" built on the Gemini family of large language models. * The agent is designed to help users interpret their IR risk predictions and provide personalized, research-backed educational content. * This AI integration aims to facilitate better communication between the data results and actionable health strategies, though it is currently intended for informational and research purposes. By utilizing ubiquitous wearable technology and existing clinical infrastructure, this approach offers a path toward proactive metabolic health monitoring. Integrating these models into consumer or clinical platforms could lower the barrier to early diabetes intervention and enable more personalized preventative care.

AILarge Language ModelsMachine LearningGemini+4

google·Aug 5, 2025Original article

Highly accurate genome polishing with DeepPolisher: Enhancing the foundation of genomic research (opens in new tab)

DeepPolisher is a deep learning-based genome assembly tool designed to correct base-level errors with high precision, significantly enhancing the accuracy of genomic research. By leveraging a Transformer architecture to analyze sequencing data, the tool reduces total assembly errors by 50% and insertion or deletion (indel) errors by 70%. This advancement is critical for creating near-perfect reference genomes, such as the Human Pangenome Reference, which are essential for identifying disease-causing variants and understanding human evolution. ## Limitations of Current Sequencing Technologies * Genome assembly relies on reading nucleotides (A, T, G, and C), but the microscopic scale of these base pairs makes accurate, large-scale sequencing difficult. * Short-read sequencing methods provide high signal strength but are limited to a few hundred nucleotides because identical DNA clusters eventually desynchronize, blending signals together. * Long-read technologies can sequence tens of thousands of nucleotides but initially suffered from high error rates (~10%); while tools like DeepConsensus have reduced this to 0.1%, further refinement is necessary for high-fidelity reference genomes. * Even a 0.1% error rate results in millions of inaccuracies across the 3-billion-nucleotide human genome, which can cause researchers to miss critical genetic markers or misidentify proteins. ## DeepPolisher Architecture and Training * DeepPolisher is an open-source pipeline adapted from the DeepConsensus model, utilizing a Transformer-based neural network. * The model was trained using a human cell line from the Personal Genomes Project that is estimated to be 99.99999% accurate, providing a "ground truth" for identifying and correcting errors. * The system takes sequenced bases, their associated quality scores, and the orientation of the DNA strands to learn complex error patterns that traditional methods might miss. * By combining sequence reads from multiple DNA molecules of the same individual, the tool iteratively "polishes" the assembly to reach the accuracy required for reference-grade data. ## Impact on Genomic Accuracy and Gene Discovery * The tool’s ability to reduce indel errors by 70% is particularly significant, as these specific errors often interfere with the identification of protein-coding genes. * DeepPolisher has already been integrated into major research efforts, including the enhancement of the Human Pangenome Reference, providing a more robust foundation for clinical diagnostics. * Improved assembly accuracy allows for better mapping of regions where the genome is highly repetitive, which were previously difficult to sequence and assemble confidently. For researchers and bioinformaticians, DeepPolisher represents a vital step in moving from "draft" genomes to high-fidelity references. Adopting this tool in assembly pipelines can drastically improve the reliability of variant calling and gene annotation, especially in complex clinical and evolutionary studies.

AIDeep LearningTransformerGenomics+4

google·Jul 31, 2025Original article

MLE-STAR: A state-of-the-art machine learning engineering agent (opens in new tab)

MLE-STAR is a state-of-the-art machine learning engineering agent designed to automate complex ML tasks by treating them as iterative code optimization challenges. Unlike previous agents that rely solely on an LLM’s internal knowledge, MLE-STAR integrates external web searches and targeted ablation studies to pinpoint and refine specific pipeline components. This approach allows the agent to achieve high-performance results, evidenced by its ability to win medals in 63% of Kaggle competitions within the MLE-Bench-Lite benchmark. ## External Knowledge and Targeted Ablation The core of MLE-STAR’s effectiveness lies in its ability to move beyond generic machine learning libraries by incorporating external research and specific performance testing. * The agent uses web search to retrieve task-specific, state-of-the-art models and approaches rather than defaulting to familiar libraries like scikit-learn. * Instead of modifying an entire script at once, the system conducts an ablation study to evaluate the impact of individual pipeline components, such as feature engineering or model selection. * By identifying which code blocks have the most significant impact on performance, the agent can focus its reasoning and optimization efforts where they are most needed. ## Iterative Refinement and Intelligent Ensembling Once the critical components are identified, MLE-STAR employs a specialized refinement process to maximize the effectiveness of the generated solution. * Targeted code blocks undergo iterative refinement based on LLM-suggested plans that incorporate feedback from prior experimental failures and successes. * The agent features a unique ensembling strategy where it proposes multiple candidate solutions and then designs its own method to merge them. * Rather than using simple validation-score voting, the agent iteratively improves the ensemble strategy itself, treating the combination of models as a distinct optimization task. ## Robustness and Safety Verification To ensure the generated code is both functional and reliable for real-world deployment, MLE-STAR incorporates three specialized diagnostic modules. * **Debugging Agent:** Automatically analyzes tracebacks and execution errors in Python scripts to provide iterative corrections. * **Data Leakage Checker:** Reviews the solution script prior to execution to ensure the model does not improperly access test dataset information during the training phase. * **Data Usage Checker:** Analyzes whether the script is utilizing all available data sources, preventing the agent from overlooking complex data formats in favor of simpler files like CSVs. By combining external grounding with a granular, component-based optimization strategy, MLE-STAR represents a significant shift in automated machine learning. For organizations looking to scale their ML workflows, such an agent suggests a future where the role of the engineer shifts from manual coding to high-level supervision of autonomous agents that can navigate the vast landscape of research and data engineering.

AILarge Language ModelsAI AgentsPython+5

google·Jul 28, 2025Original article

Simulating large systems with Regression Language Models (opens in new tab)

Researchers from Google have introduced Regression Language Models (RLMs) as a universal solution for numeric prediction tasks by framing regression as a text-to-text problem. By converting complex, unstructured system data into strings, RLMs can predict performance metrics without the need for manual feature engineering or data normalization. This approach allows large language models to move beyond subjective human feedback and directly model raw operational data for large-scale software and industrial infrastructures. ## Conceptualizing Text-to-Text Regression * Traditional regression methods rely on tabular data—fixed-length numeric vectors—which are difficult and laborious to maintain for evolving systems like software logs or hardware patterns. * RLMs represent the input state ($x$) as a structured text string (such as JSON or YAML) and the numerical output ($y$) as a text string. * The model is trained using standard next-token prediction and cross-entropy loss, allowing it to function as a universal approximator for complex data types. * This paradigm eliminates the need for manual feature engineering, as the model learns directly from the raw textual representation of the system state. ## Architecture and Training for Large Systems * The research utilizes a compact RLM consisting of a two-layer encoder-decoder architecture with 60 million parameters. * To manage large inputs that can reach up to 1 million tokens, the system reorders features by importance at the beginning of the string so that critical data is preserved when truncated to the model's 8k token limit. * Pre-training the RLM on diverse regression tasks enables few-shot adaptation, allowing the model to adjust to new data types with minimal gradient updates. * Numerical values are processed as-is within the text, removing the requirement for traditional scaling or normalization common in standard machine learning pipelines. ## Optimizing Google's Borg Infrastructure * The method was specifically applied to Google’s Borg system to predict MIPS per GCU (Millions of Instructions Per Second per Google Compute Unit), a vital efficiency metric. * The RLM simulates the outcomes of complex bin-packing algorithms within a "digital twin" framework to optimize resource allocation across CPUs and TPUs. * By analyzing execution traces and textual metadata, the model provides high-accuracy forecasting for diverse workloads including Gmail, YouTube, and Maps. ## Density Capture and Uncertainty Modeling * Unlike traditional regressors that provide a single point estimate, RLMs can capture full probability distributions by sampling the decoded output multiple times. * This density estimation is critical for modeling aleatoric uncertainty, which represents the inherent randomness and stochastic load demands of large-scale compute environments. * The ability to visualize these distributions helps engineers identify the range of possible outcomes and the inherent variability of the system's performance over time. This research demonstrates that small, specialized language models can effectively replace traditional regression methods in highly dynamic environments. For practitioners looking to implement these capabilities, the open-source `regress-lm` library provides a framework for simulating large systems and predicting performance across varied industrial and scientific use cases.

AILarge Language ModelsGoogle BorgRegression Language Models+5

google·Jul 27, 2025Original article

SensorLM: Learning the language of wearable sensors (opens in new tab)

SensorLM is a new family of foundation models designed to bridge the gap between high-dimensional wearable sensor data and natural language descriptions. By training on a massive dataset of nearly 60 million hours of de-identified health data, the models learn to interpret complex physiological signals to provide meaningful context for human activities. This research demonstrates that integrating multimodal sensor signals with language models enables sophisticated health insights, such as zero-shot activity recognition and automated health captioning, that significantly outperform general-purpose large language models. ## Dataset Scale and Automated Annotation * The models were pre-trained on an unprecedented 59.7 million hours of multimodal sensor data collected from over 103,000 individuals across 127 countries. * To overcome the high cost of manual annotation, researchers developed a hierarchical pipeline that automatically generates text descriptions by calculating statistics and identifying trends within the raw sensor streams. * Data was sourced from Fitbit and Pixel Watch devices, representing nearly 2.5 million person-days of activity and health information. ## Hybrid Training Architecture * SensorLM unifies two primary multimodal strategies: contrastive learning and generative pre-training. * Through contrastive learning, the model learns to discriminate between different states—such as a "light swim" versus a "strength workout"—by matching sensor segments to corresponding text descriptions. * The generative component allows the model to "speak" for the sensors, producing nuanced, context-aware natural language captions directly from high-dimensional biometric signals. ## Activity Recognition and Cross-Modal Capabilities * The model demonstrates state-of-the-art performance in zero-shot human activity recognition, accurately classifying 20 different activities without any specific fine-tuning. * Its few-shot learning capabilities allow the model to adapt to new tasks or individual user patterns with only a handful of examples. * SensorLM facilitates cross-modal retrieval, enabling users or experts to find specific sensor patterns using natural language queries or to generate descriptions based on specific sensor inputs. ## Generative Health Captioning * Beyond simple classification, the model can generate hierarchical captions that describe the statistical, structural, and semantic dimensions of a user’s data. * Experimental results using metrics like BERTScore show that SensorLM produces captions that are more factually correct and coherent than those created by powerful non-specialist LLMs. * This capability allows for the translation of abstract data points, such as heart rate variability or step counts, into readable summaries that explain the "why" behind physiological changes. By providing a framework where wearable data can be understood through the lens of human language, SensorLM paves the way for more intuitive and personalized health monitoring. This technology holds the potential to transform raw biometric streams into actionable insights, helping users better understand the relationship between their activities and their overall physical well-being.

AIMultimodal AIFoundation ModelsWearable Technology+5

google·Jul 23, 2025Original article

Synthetic and federated: Privacy-preserving domain adaptation with LLMs for mobile applications (opens in new tab)

Researchers at Google have developed a framework for improving both small and large language models (LMs) in mobile applications like Gboard by utilizing privacy-preserving synthetic data and federated learning. This approach combines differential privacy (DP) with large language model (LLM) generation to minimize data memorization risks while achieving significant gains in production metrics like next-word prediction and proofreading. The result is a robust pipeline that allows models to adapt to specific user domains without compromising individual privacy or requiring centralized data storage. ### Strengthening Privacy with DP-FL * Gboard has transitioned all production LMs trained on user data to a Federated Learning with Differential Privacy (DP-FL) framework, ensuring data remains on-device and is never memorized. * The deployment utilizes the **BLT-DP-FTRL** algorithm, which offers an optimized trade-off between privacy guarantees and model utility while being easier to deploy in production. * Engineers adopted the **SI-CIFG** model architecture to facilitate efficient on-device training, ensuring the hardware can handle local updates while maintaining compatibility with DP constraints. ### Synthetic Data Generation via Public LLMs * Powerful LLMs trained on public web data are prompted to synthesize high-quality text that mimics mobile user interactions without ever accessing actual private user data. * The process involves a two-step prompting strategy: first, filtering public datasets to identify topics common in mobile communication, and second, generating new, domain-specific text based on those patterns. * This synthetic data serves as a bridge for pre-training small LMs, which are then refined through private post-training on-device to capture the nuances of user behavior. ### Adapting LLMs for Mobile Proofreading * To support advanced features like Gboard's "Proofread," researchers developed a "Synthesize-then-Adapt" pipeline specifically for error correction. * LLMs generate synthetic "corrupted" text to simulate common mobile typing errors, providing the necessary training pairs (error/correction) that are difficult to find in public datasets. * Federated learning is then used to adapt these error-correction models to specific app domains (such as messaging or email) using on-device signals, ensuring the model understands the specific context of the user's typing. The success of these techniques in Gboard demonstrates that synthetic data can effectively replace or augment private data throughout the machine learning lifecycle. For developers working with sensitive user information, adopting a "synthetic-first" approach combined with federated learning provides a scalable path to model improvement that adheres to the core principles of data minimization and anonymization.

Large Language ModelsDifferential PrivacySynthetic DataOn Device Ai+3