Google Research

In 2023, Stanford University and Google Research jointly published the paper "Generative Agents: Interactive Simulacra of Human Behavior," deploying 25 LLM-based agents in a virtual town.

思维链（Chain-of-Thought）提示技术由Google Research于2022年提出

Self-Consistency was formally proposed by Google Research in the 2022 paper 'Self-Consistency Improves Chain of Thought Reasoning in Language Models.'

The concept of emergent abilities was systematically elaborated by the Google Research team (Wei et al.) in the 2022 paper Emergent Abilities of Large Language Models, which examined more than 200 tasks.

In a 2022 study, Google Research systematically summarized the concept of emergent capabilities, finding that abilities such as chain-of-thought reasoning and few-shot learning are absent in small models but suddenly emerge after the parameter count crosses a critical threshold.

SigLIP is a vision-language contrastive learning model proposed by Google Research that uses a sigmoid loss function to replace CLIP's softmax contrastive loss, eliminating the dependency on global negative samples.

xAI's technical team brings together researchers from top institutions including DeepMind, OpenAI, and Google Research.

Titans (Google Research) introduces a neural long-term memory module that updates memory based on surprise at test time.

Yossi Matias is the Vice President and Head of Google Research.

Google Research proposed the RRSI (Regularized Recursive Self-Improvement) framework for constraining the recursive self-improvement process of AI agents

JAX is developed by Google Research, and its core strengths lie in its functional programming style and the composability of jit, vmap, and grad.

Data Commons was built by Google Research starting around 2018.

FreshQA is an evaluation benchmark released by Google Research in 2023, used to test the ability of language models and search systems to handle time-sensitive questions.

In 2022, researchers from Google Research and Stanford systematically described the phenomenon of emergent abilities for the first time in a paper: when model scale crosses a certain threshold, performance on specific tasks jumps from near-random to significantly above baseline.

According to a Google Research paper, the pre-training data scale of TimesFM reaches approximately 100 billion time points.

The GQA technique was initially proposed by Google Research in 2023 and was subsequently widely adopted by mainstream models such as Llama 2 and Mistral.

Some top industrial laboratories (such as Google Research, Meta AI, and Microsoft Research) regard Findings papers as nearly equivalent to main conference papers, while certain traditional academic institutions assign them lower weight in faculty review.

The NIMA model released by Google Research in 2017 uses MobileNet as its backbone network, was trained on the AVA dataset containing more than 250,000 photos, and can predict the distribution of human aesthetic scores.

The supervised contrastive learning framework SupCon was proposed by Google Research in 2020 and is a supervised extension of self-supervised contrastive learning.

Google Research released TabFM, a foundation model built specifically for tabular data.