Olivier Bachem

Olivier Bachem is a machine learning researcher at Google DeepMind who works on the finetuning and open release of large language models. He studied economics, quantitative finance and statistics before receiving a PhD at ETH Zurich under Andreas Krause, where he worked on coresets for large-scale clustering and became a 2016 Google PhD Fellow. In 2019, his paper with Francesco Locatello and colleagues on disentangled representations won an ICML Outstanding Paper award. At Google, he coauthored the Gemini, Gemma, Gemma 2, Gemma 3 and Gemini 2.5 reports and helped build the initial reinforcement learning approach used in Gemini. He is a Research Scientist Senior Director at Google DeepMind, where he co-leads the Gemma program for Google's open-weight and on-device models.

Now at: Google DeepMind

Timeline

By 2016

A Google PhD Fellowship

He earns a bachelor's degree in economics at the University of St. Gallen and master's degrees in quantitative finance (ETH Zurich and the University of Zurich) and statistics (ETH Zurich), where his master thesis wins an ETH medal. He receives his PhD from ETH Zurich under Andreas Krause in the Learning & Adaptive Systems group, and Google names him a 2016 Google PhD Fellow in Machine Learning.
2018

Lightweight coresets for clustering

In 2018, he coauthors Scalable k-Means Clustering via Lightweight Coresets with Mario Lučić and Andreas Krause for the International Conference on Knowledge Discovery and Data Mining; the construction is substantially faster than existing ones and yields smaller coresets. His dissertation investigates coresets, small summaries of large data sets with theoretical guarantees, and other sampling methods for large-scale clustering.
2019

An ICML Outstanding Paper

In 2019, Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations, written with Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Rätsch, Sylvain Gelly and Bernhard Schölkopf, receives an Outstanding Paper award at ICML. The paper shows theoretically that unsupervised learning of disentangled representations is impossible without inductive biases, and trains more than 12,000 models in a large-scale experimental study.
2020–2022

Large-scale reinforcement learning studies

In 2020, he coauthors What Matters In On-Policy Reinforcement Learning?, a study that trains over 250,000 agents and becomes an oral presentation at ICLR 2021. His 2022 paper A general class of surrogate functions for stable and efficient reinforcement learning earns a Best Paper Award Honorable Mention at AISTATS 2022, and Concave Utility Reinforcement Learning: the Mean-Field Game Viewpoint is Best Paper Runner-up at AAMAS 2022.
2023

Finetuning behind Gemini

In December 2023, he is among the authors of the Gemini report, which introduces the Ultra, Pro and Nano multimodal models. He helps build finetuning techniques used in various Google projects, including the initial reinforcement learning approach used in Gemini.
2024

Gemma and Gemma 2

In March 2024, he is among the authors of the Gemma technical report, which introduces open models built from Gemini research and technology in 2 billion and 7 billion parameter sizes. The Gemma 2 report, with him again among its authors, follows in July 2024 with models from 2 billion to 27 billion parameters.
2025

Gemma 3 and Gemini 2.5

The Gemma 3 report, posted in March 2025 with him among its authors, covers multimodal models from 1 to 27 billion parameters with vision understanding, wider language coverage and context of at least 128K tokens. In July 2025, he is also among the authors of the Gemini 2.5 report, which introduces Gemini 2.5 Pro and Gemini 2.5 Flash.
As of October 2026

Co-leading the Gemma program

He is a Research Scientist Senior Director at Google DeepMind, co-leading the Gemma program, which produces Google's open-weight and on-device LLMs. The Gemma 4 technical report, posted in July 2026 with him among its authors, describes a model suite with dense and mixture-of-experts architectures from 2.3B to 31B parameters.