Research papers, technical reports, benchmarks and datasets.
A paper from Google Research at COLM 2026 finds that reasoning traces can unlock parametric knowledge in LLMs that they could not otherwise access, even for simple single-hop factual questions. Tests on Gemini-2.5 Flash and Pro and Qwen3-32B identify two mechanisms: extra tokens act as a “computation buffer,” and “fact priming” builds a retrieval bridge by generating related facts.
Researchers at Oxford, the UK's AI Safety Institute, Stanford, and the London School of Economics ran four experiments with 6,923 participants and 18,978 conversations. They found that AI systems consistently outperformed human experts at text-based persuasion, even when experts chose their own topics, researched in advance, and were offered a £1,000 prize incentive.
Google Research has released a vectorized dataset that converts the high-resolution raster maps in Farmscapes 2020 into actionable inventories of hedgerows, stone walls, and scrubland, covering more than 130,000 square kilometers of the UK. The framework fine-tunes an RSF Vision Transformer backbone pretrained on over 300 million global satellite images and uses Polsby–Popper compactness scores to distinguish woodland, woody patches, and linear woody features.
Google reports on consumer understanding of dermatology AI tools. Among 2,345 participants, more than 62% of those using AI tried to name a skin condition, with 23% accuracy—nearly three times the control group's 8%. The “Wizard of Oz” group reached 36% accuracy. But the AI group showed no statistically significant improvement in deciding what care to seek and was slightly more likely than dermatologists to recommend less urgent care (30% vs. 27%).
At AISTATS 2026, Google Research proposed Regularized f-Divergence Kernel Tests, a relative-distance test for determining whether an unlearned model's distribution is closer to a safe retrained model or to the original compromised model.
Google DeepMind has published results from a preregistered randomized controlled trial in Sierra Leone. Students using Guided Learning improved their math scores by 0.258 standard deviations over the control group—equivalent to about 1.2 to 1.7 years of typical learning—in eight weeks.
Why it matters: The study reports quantified results and interaction data from a randomized trial in Sierra Leone, helping readers assess the effects of AI-assisted instruction in real classrooms.
In a paper published in Nature, Google Research describes PHRM, a system that uses the front-facing camera to record a few seconds of video after face unlock and estimate heart rate and resting heart rate in the background using deep learning.
Why it matters: Google passively measures heart rate with a phone's front camera after unlock and has released its largest phone-video dataset to date along with a pretrained model.
Google has introduced zero-trust aggregation for its privacy analytics service, combining a new cryptographic aggregation method with hardware-protected TEEs. It can receive device data in a single message without multi-round online interaction. Google receives only anonymous, aggregate insights about groups; raw data is neither exposed nor reconstructible, even within the hardware-protected environment. TEE attestation provides participants with proof that the protocol ran as intended.
Google has announced that Empirical Research Assistance (ERA), a Gemini-based tool for scientific coding, is available to scientists. The related work has been published in Nature.
Why it matters: Google has moved ERA from a Nature paper to an application-based tool, offering a view of the practical limits of AI-written scientific code.
Google DeepMind biologists Omar Abudayyeh and Jonathan Gootenberg used Co-Scientist to support aging research. After scanning tens of thousands of papers, the tool proposed more than 20 new candidate genetic factors. Lab tests validated several hypotheses, successfully making cells younger and improving their overall function. Co-Scientist also analyzed large-scale genetic-screening data, reducing data and literature linkage work that could take up to six months to just days.
Why it matters: The source reports specific outputs and time comparisons from Co-Scientist's role in aging research, showing how AI agents contribute to scientific hypothesis generation.
Matt Onsum and Katherine Labbé of Calico Life Sciences are using Google DeepMind's Co-Scientist to integrate scattered findings in aging biology and generate testable hypotheses.
University of Edinburgh bioengineer Filippo Menolascina used Google Co-Scientist to study MASH liver disease. The system integrated evidence from liver biology and pharmacology, identified mechanisms worth investigating, and flagged candidate combination therapies for testing.
Google's Co-Scientist helped MIT mechanical engineer Ritu Raman rapidly review ALS literature, turn ideas into testable hypotheses, and rank them by feasibility. Its strongest lead pointed to interactions between molecules on cell surfaces, outside Raman's field, prompting her to work with Boston Children's Hospital chemical biologist Ryan Flynn. Together, using their respective toolkits and iterating with Co-Scientist, they are searching for new RNA mechanisms—and potentially RNA drugs—that could target ALS.
Stanford School of Medicine geneticist Gary Peltz used Google DeepMind's Co-Scientist to screen for anti-fibrotic drugs. Of the three candidates it proposed, two successfully blocked fibrosis and promoted liver-cell regeneration in experiments with live human liver cells; neither of the two candidates Peltz selected himself was effective.
Google DeepMind has announced an AI co-clinician research program exploring a “triple-care” model in which AI agents assist with patient care under clinical supervision. In blind evaluation of 98 real primary-care queries, the system had no critical errors in 97 cases, and doctors preferred its responses to those from existing evidence-synthesis tools. It also outperformed other frontier models on open-ended medication questions in OpenFDA's RxQA.
Why it matters: Google DeepMind reports results from physician blind reviews, RxQA, and simulated remote consultations, offering a view of current medical-AI capabilities.
Google DeepMind has released Decoupled DiLoCo, a distributed-training architecture that divides large-scale training into compute islands communicating asynchronously, reducing bandwidth requirements and isolating hardware failures. In tests, training continued and reconnected after an entire learner unit went offline, achieving the same ML performance as traditional methods with Gemma 4 models.
Why it matters: The source reports distributed pretraining results across bandwidth, fault tolerance, and mixed-generation hardware, helping readers assess engineering tradeoffs.
In an ICLR paper, Google Research proposes ReasoningBank, an agent-memory framework that extracts high-level, structured reasoning memories from successful and failed experiences. Its code is open-source.
Why it matters: The paper describes a closed loop for extracting reasoning memories from successes and failures, with success-rate and step-count results on WebArena and SWE-Bench-Verified.
A Berkeley AI Research team has proposed GRASP, a gradient-based planner that makes long-horizon planning with learned dynamics (world models) feasible. It lifts trajectories into virtual states for parallel optimization across time, injects stochasticity directly during iterative state refinement to explore, and reshapes gradients so actions receive clearer signals. The method addresses ill-conditioned computation graphs, local minima caused by non-greedy structure, and fragile gradients in high-dimensional visual models.
Google Research has published “Reasoning-Driven Synthetic Data Generation and Evaluation,” proposing Simula, a framework that treats synthetic-data generation as a mechanism-design problem. A reasoning model builds an entire dataset from scratch without seed data.
Google Research has proposed MoGen (Neuronal Morphology Generation), a model that generates synthetic neuron shapes with point-cloud flow matching. The data supplements training for the PATHFINDER reconstruction model, reducing mouse-axon reconstruction error by 4.4%, mainly by decreasing merge errors. At whole-mouse-brain scale, the work is equivalent to saving 157 person-years of manual proofreading. MoGen is available as an open-source model.
Jan Leike reports that Claude makes fully autonomous progress on scalable oversight research, measured by performance gap recovered (PGR). Claude iterates on several different techniques and ends up significantly outperforming human researchers for $18k in credits.
Google has released Vantage as a research experiment. It uses generative AI to create simulated conversations that assess high-school and university students' future skills, including problem-solving and collaboration. It is available for English registration through Google Labs. An Executive LLM dynamically guides the conversation and an AI Evaluator scores it against a rubric. In a study with New York University involving 188 participants aged 18–25, AI scores aligned with human experts nearly as well as experts aligned with one another.
Google Research has introduced ConvApparel, a fashion-shopping dataset containing more than 4,000 multi-turn human–AI conversations (nearly 15,000 turns) for measuring the realism gap between LLM user simulators and real human behavior.
Google has released two AI agents for academic workflows: PaperVizAgent generates figures for papers, while ScholarPeer automates peer review. PaperVizAgent coordinates five agents for retrieval, planning, style, visualization, and review. It scored 60.2 on a 0–100 evaluation, ahead of GPT-Image-1.5, Nano-Banana-Pro, and Paper2Any, and was the only framework to exceed the human baseline of 50.0.
Google Research proposes a framework using situational judgment tests (SJTs), adapted from psychology scales, to assess how closely LLM behavioral tendencies align with human consensus across four traits including empathy and emotional regulation. A large-scale analysis of 25 LLMs finds two kinds of gap: models' tendencies can diverge from annotator consensus, and models may fail to cover the range of human opinions when there is no consensus. Small models under 120B and frontier closed models align nearly perfectly when humans unanimously agree, but alignment levels off just above 80% when consensus is below 90%.
In the paper “Forest vs Tree: The (N, K) Trade-off in Reproducible ML Evaluation,” Google Research studies the tradeoff between the number of evaluation items and annotators per item. It finds that the commonly used one to five annotators per item is often insufficient; more than ten may be needed to capture variation in human opinion.
Google Quantum AI has published a white paper saying future quantum computers may be able to break the elliptic-curve cryptography protecting cryptocurrencies with fewer qubits and gates than previously thought.
Why it matters: Google Quantum AI presents new estimates of the quantum resources needed to break ECDLP-256 and explains how it disclosed the vulnerability without revealing attack details.
Google has released TurboQuant, a compression algorithm that quantizes KV cache to 3 bits without training or fine-tuning and without loss of model accuracy, reducing KV-memory use by at least 6×.
Why it matters: Google reports TurboQuant's algorithm details and benchmark results, helping readers assess the usefulness of KV-cache compression for long-context workloads.
Google Research has released S2Vec, a self-supervised framework that converts the built environment into layered images through S2 Geometry tiling and feature rasterization, then learns general-purpose embeddings with a masked autoencoder (MAE). On zero-shot geographic extrapolation for U.S.-wide socioeconomic predictions such as population density and median income, S2Vec is usually the best single model. Multimodal fusion with image embeddings generally performs better than either modality alone, though tasks such as tree canopy and elevation remain challenging.
Two companion studies in Nature Cancer evaluate an AI breast-screening system in the AIMS research collaboration between Google Research and several NHS organizations.
Why it matters: The two Nature Cancer studies report detection-rate and workload results for AI in the NHS double-reading workflow, showing the practical limits of human–AI collaboration.
Google and Cornell University tested GPT-4o, Perplexity, Claude 3.5, Gemini Advanced Pro 1.5, NotebookLM, and a custom RAG system on 67 expert-level questions about high-temperature superconductivity. An international panel of experts scored the answers blind across six dimensions.
Berkeley AI Research has proposed SPEX and ProxySPEX, algorithms that identify important interactions driving LLM outputs across feature, data, and model-component attribution. SPEX turns interaction search into a sparse-recovery problem by exploiting sparsity and low-order structure; ProxySPEX uses additional hierarchy to match SPEX performance with about ten times fewer ablation experiments.
Google Research and Beth Israel Deaconess Medical Center conducted a prospective, single-center feasibility study in which AMIE interviewed patients by text before outpatient primary-care visits, under video supervision by a physician throughout.
Why it matters: Google and BIDMC have deployed conversational diagnostic AI in a real outpatient workflow for the first time, showing its safety-supervision model and diagnostic performance.
A Berkeley AI Research team proposes a mutual-information-based framework for evaluating and optimizing imaging systems. It estimates information directly from noisy measurements to predict system performance. Validated in color photography, radio astronomy, lensless imaging, and microscopy, optimized designs reach end-to-end-method performance with lower memory and compute needs and no task-specific decoder. The paper appeared at NeurIPS 2025.
Berkeley AI Research has proposed a divide-and-conquer off-policy reinforcement-learning algorithm that does not rely on temporal-difference (TD) learning. It reduces the number of Bellman-recursion steps from linear to logarithmic, enabling scaling to long-horizon tasks. The method achieves scaling for the first time in goal-conditioned RL, using the triangle inequality for shortest-path distances between states to construct transitive Bellman updates. Selecting the optimal subgoal w remains an open problem.
Berkeley AI Research proves that under a fourth-order approximate objective, specific hyperparameter constraints, tiny initial embedding weights, and tiny gradient-descent steps, word2vec's learning problem reduces to an unweighted least-squares matrix factorization. Its gradient-flow dynamics have a closed-form solution, and the resulting representations are equivalent to applying PCA to matrix M*.
