Max Bartolo

Max Bartolo is a researcher whose work has tested how language models handle adversarial examples, human feedback and tool use. He joined Bloomsbury AI in 2017, began a PhD in the UCL NLP group in 2019, and studied adversarial robustness with Facebook AI Research. At Cohere, he built and led the post-training team behind Command Nightly, Command R+ and Command A; his research on few-shot prompting won an ACL Outstanding Paper award in 2022. His coauthored PRISM dataset brought together stated preferences from 1,500 participants in 75 countries and feedback from 8,011 conversations with 21 language models, and won NeurIPS’s 2024 Datasets and Benchmarks Best Paper. Since August 2025, he is a researcher at Google DeepMind working on Gemini’s agentic robustness, tool use and function calling.

Now at: Google DeepMind
Badges: Bridge

Timeline

2017–2018

Question answering at Bloomsbury AI

He joins Bloomsbury AI in May 2017 to work on open-domain question answering. In 2018, he coauthors the EMNLP paper Interpretation of Natural Language Rules in Conversational Machine Reading, which formalises a task that asks models to combine written rules with background knowledge to answer questions.
2019

A PhD at UCL

In January 2019, he starts a PhD in the UCL NLP group under Pontus Stenetorp and Sebastian Riedel, studying the adversarial robustness of language models with humans and models in the loop.
2020–2021

Adversarial work at Facebook AI Research

In September 2020, he joins Facebook AI Research as a research intern on adversarial benchmarking and robustness. His 2021 EMNLP paper with Maximilian Mozes and colleagues compares human-generated and machine-generated word-level adversarial examples for text classification.
2021–2022

Annotation assistants and prompt order

In 2021, he coauthors Models in the Loop, which gives crowdworkers real-time suggestions from generative annotation assistants. His Fantastically Ordered Prompts paper uses synthetic examples to rank few-shot prompt orders and wins an ACL Outstanding Paper award in 2022.
October 2022–August 2025

Cohere post-training team

In October 2022, he joins Cohere to make language models more useful and robust. Before joining Google DeepMind in August 2025, he builds and leads Cohere’s post-training team, shipping Command Nightly, Command R+ and Command A. In 2024, he coauthors PRISM, a dataset of stated preferences from 1,500 participants in 75 countries and fine-grained feedback from 8,011 live conversations with 21 language models; the paper wins NeurIPS’s Datasets and Benchmarks Best Paper.
Since August 2025

Gemini agent robustness

He joins Google DeepMind in August 2025 to work on reliable, robust function calling and tool use for Gemini. He is a researcher there, working on Gemini’s agentic robustness, tool use and function calling.