Google Research
02 Aug 2023
Researchers from Google Brain and Google Research introduce the Transformer, a novel neural network architecture that replaces recurrent and convolutional layers entirely with attention mechanisms. The Transformer achieves new state-of-the-art performance on machine translation benchmarks, securing a BLEU score of 28.4 on WMT 2014 English-to-German, while drastically reducing training costs and demonstrating strong generalization to other NLP tasks like constituency parsing.
03 Jun 2021
Google Research's Vision Transformer (ViT) demonstrates that a pure Transformer architecture can achieve state-of-the-art performance in image classification by treating images as sequences of patches. When pre-trained on large datasets such as JFT-300M, ViT models surpass leading CNNs like BiT and EfficientNet across multiple benchmarks while requiring substantially fewer computational resources for pre-training.
25 Oct 2025
Mixture-of-Recursions (MoR) introduces a unified framework for language models that combines parameter efficiency, adaptive computation, and efficient KV caching. It achieves strong performance, outperforming vanilla Transformers with nearly 50% fewer parameters and reducing training FLOPs by 25%, while increasing inference throughput by up to 2.06x.
31 Dec 2024
Titans introduces a new neural memory module that learns to memorize information at test time and effectively combines with attention mechanisms for better long-term sequence modeling
28 May 2024
Researchers at Google Research developed a method to convert one-sided differentially private (DP) algorithms into two-sided pure-DP algorithms through subsampling. This framework yields improved or asymptotically tight error bounds for pure-DP solutions in various combinatorial optimization and streaming problems, including submodular maximization, set cover, k-means/k-median, and shifting heavy hitters.
06 Aug 2023
Researchers at Stanford University and Google present 'generative agents,' an architecture that extends large language models to create interactive simulacra of human behavior. The system incorporates dynamic memory, recursive reflection, and hierarchical planning, enabling individual agents to act believably and fostering emergent social dynamics like information diffusion and relationship formation in a simulated environment.
26 Jul 2022
Google Research's Brain Team introduces classifier-free guidance, a technique for diffusion models that enables high-quality image generation without a separate classifier network. This method involves training a single diffusion model to function in both conditional and unconditional modes, providing a simple parameter to control the trade-off between image quality and diversity.
19 May 2023
Google Research developed Speculative Decoding to accelerate large Transformer model inference without altering the output distribution or requiring model retraining. The method uses a smaller model to speculatively guess future tokens, which are then verified in parallel by the larger model, achieving 2X-3X walltime speedups on tasks like translation and summarization.
21 Jul 2025
Researchers at Google Research developed a theoretical framework that explains In-Context Learning (ICL) in Large Language Models as an implicit, low-rank modification of the Multi-Layer Perceptron (MLP) layer's weights, rather than just the self-attention mechanism. Their model provides explicit formulas for these implicit weight updates, and experiments demonstrate a strong correspondence between the predicted implicit updates and actual ICL behavior in a transformer.
25 Aug 2022
A Google Research paper unifies different theoretical perspectives on diffusion models, showing they can be understood from both likelihood-based (variational inference) and score-based approaches. It establishes that diffusion models are a specialized form of Hierarchical Variational Autoencoders (HVAEs), demonstrating the mathematical equivalence of their three primary training objectives.
11 Aug 2023
Google's RT-1 is a Transformer-based model trained on 130,000 real-world robotic demonstrations across 700 language-conditioned tasks, achieving a 76% success rate on unseen instructions and exhibiting strong robustness to distractors and novel environments. It effectively integrates diverse data sources, including simulation and data from other robot morphologies, with minimal performance degradation on original tasks.
25 May 2023
SELF-REFINE enables a single large language model to iteratively refine its own outputs by generating self-feedback and then using that feedback to improve the generation. This approach led to an average absolute performance improvement of approximately 20% across seven diverse generative tasks, consistently outperforming one-step generation and a multiple-sample baseline.
14 May 2025
Google DeepMind
University of Illinois at Urbana-ChampaignUniversity of Freiburg
Carnegie Mellon University
Imperial College London
University of Southern California
New York University
Shanghai Jiao Tong University
the University of Tokyo
Stanford University
The University of Texas at AustinUniversity of Technology Nuremberg
ETH Zürich
University of California, San Diego
RIKENGoogle Research
Columbia University
Arizona State UniversityurichGerman Aerospace CenterIstituto Italiano di TecnologiaMax Planck InstituteQueensland University of Technologyat DarmstadtKorea Advanced Institute of Science & TechnologyIntrinsic LLCFlexiv RoboticsTechnische Universit
The OpenX-Embodiment Collaboration released the Open X-Embodiment (OXE) Dataset, a consolidated collection of over 1 million real robot trajectories from 22 embodiments. This work demonstrates that large RT-X models trained on such diverse data achieve positive transfer and emergent skills across different robot platforms.
26 Feb 2025
A groundbreaking multi-institutional collaboration introduces an AI co-scientist system built on Gemini 2.0 that successfully generates and validates novel scientific hypotheses across biomedical domains, achieving experimental validation of drug repurposing predictions and independently discovering unpublished findings in bacterial evolution.
17 Apr 2024
Google Research developed TimesFM, a decoder-only foundation model for time-series forecasting, which achieves strong zero-shot performance across diverse, unseen datasets. The model, pre-trained on a massive corpus of real and synthetic time series, demonstrates that a purpose-built model can outperform general large language models repurposed for time-series tasks.
12 Jul 2025
Google Research and DeepMind introduced MedGemma, a collection of open, medically-tuned vision-language foundation models built upon Gemma 3. These models demonstrate advanced medical understanding and reasoning, outperforming similar-sized general-purpose models on medical tasks while retaining general capabilities.
08 Feb 2024
Matryoshka Representation Learning from researchers at the University of Washington, Google Research, and Harvard University develops a method to encode multi-fidelity information within a single embedding, allowing truncated prefixes to serve as progressively finer-grained representations. This approach enables substantial efficiency gains in large-scale classification and retrieval tasks, achieving comparable accuracy to full-dimensional models with significantly reduced computational cost and memory footprint.
23 Dec 2023
Google Research introduces Grouped-Query Attention (GQA) and an "uptraining" method to convert existing Multi-Head Attention (MHA) models into more efficient GQA/MQA architectures. This approach allows for substantial inference speedups, up to 5.4x, with minimal quality degradation and significantly reduced computational cost (5% of original pre-training) for large language models.
17 Apr 2025
A unified framework from Google Research introduces "Miras" for sequence modeling that reinterprets attention mechanisms and forgetting gates as forms of memory optimization, leading to three new architectures (Moneta, Yaad, and Memora) that demonstrate improved performance on language modeling and recall-intensive tasks while providing theoretical insights into existing models.
11 Sep 2020
This paper from Google Research proposes EfficientNet, a family of convolutional neural networks that achieve state-of-the-art accuracy with significantly fewer parameters and FLOPs. It introduces a novel compound scaling method that uniformly scales network depth, width, and resolution using a single set of coefficients, demonstrating that balanced scaling is crucial for optimizing performance and efficiency.
There are no more papers matching your filters at the moment.
