Papers
arXiv 2026
Combines formal specifications with gradient-based learning so self-evolving LLM agents improve while provably meeting their contracts.
arXiv 2025
Treats LLM workflows as typed probabilistic programs with parameter-efficient adapters, so the whole workflow can be trained end-to-end with gradients instead of prompt tuning.
Technical report, 2025
The Gemini 2.X model family, with state-of-the-art coding and reasoning, long video understanding and new agentic workflows.
NeurIPS 2024
Introduces the LongFact benchmark, SAFE (a search-augmented automatic fact checker) and the F1@K metric for measuring factuality in long answers.
arXiv 2024
A survey of how synthetic data is generated and used to train and evaluate language models, and how to keep it factual, faithful and unbiased.
arXiv 2024
Introduces MoLA, which allocates different numbers of LoRA experts to different Transformer layers, showing that more experts in higher layers gives better results with fewer parameters.
TMLR 2024
Trains language models to do numeric regression from text alone, beating conventional regressors when trained across many tasks.
Technical report, 2023
The first Gemini models, natively multimodal across image, audio, video and text, with Gemini Ultra setting the state of the art on 30 of 32 benchmarks.
ICML 2023
A non-uniform Transformer architecture found by evolutionary search over layer types and sparsity, training faster and scaling better than hand-designed models like GLaM.
arXiv 2023
Searches architectures layer by layer, turning multi-objective NAS into a problem of polynomial complexity while still finding better models.
arXiv 2023
Uses symbolic patches so researchers can share ML ideas as rules that apply to models they have never seen, cutting code across teams.
CVPR 2022
Fuses lidar and camera features deep in the network, using InverseAug and LearnableAlign to keep them aligned, for better 3D detection in self-driving.
AutoML Conference 2022
Shows that DARTS finds better policy networks for reinforcement learning across algorithms and environments, at about 3x the cost of training one model.
arXiv 2021
Meta-learns RL algorithms by searching over computational graphs of loss functions, rediscovering temporal-difference learning and finding variants that improve on DQN.
arXiv 2021
Jointly searches neural architectures and edge accelerator configurations, improving ImageNet accuracy by about 1% or cutting energy by up to 2x.
arXiv 2021
Combines evolution strategies with combinatorial optimizers to search spaces that mix discrete and continuous parameters.
OpenReview 2021
Searches neural architectures and hardware accelerator designs together rather than separately, for better accuracy and latency on edge devices.
arXiv 2020
A single pipeline that searches architectures and hyperparameters together, alternating weight training with an RL controller.
arXiv 2020
Uses Neural Network Gaussian Process inference as a cheap signal for ranking architectures during neural architecture search.
NeurIPS 2020 (oral)
A symbolic programming library that decouples search spaces, search algorithms and child programs, so AutoML ideas can be changed without rewriting program logic.
arXiv 2018
Shows that choosing pre-training data matters for transfer, and reweights it by importance to the target domain to improve fine-grained classification.
Patents
US Patent 12,443,882 B2, 2025 (continuation US 2026/0105369 A1)
US 2025/0238683 A1 (application), 2025
US 2024/0289605 A1 (application), 2024
US 2024/0112027 A1 (application), 2024
US 2024/0005129 A1 (application), 2024
US 2022/0391687 A1 (application), 2022
US 2022/0019856 A1 (application), 2022
US 2021/0383223 A1 (application), 2021
US 2020/0104710 A1 (application), 2020