Senior Research Scientist · Google DeepMind

Roma Patel

I am a Senior Research Scientist at Google DeepMind working on AI safety and alignment , with a particular interest in how increasingly capable language-model systems can be evaluated, aligned, and stress-tested before their failures become consequential.

My research asks how we can understand failures in language models, improve alignment objectives and training methods, and build systems that proactively surface emerging risks. My earlier work, including my PhD at Brown University with Ellie Pavlick, studied language grounding, representation, and interpretability: foundations that continue to shape how I think about what models understand and how their behavior can be made legible.

I am broadly interested in the intersection of language, reinforcement learning, multi-agent systems, and long-term AI safety. Outside of work, I enjoy reading vast amounts of literature, going to new mountains, and finding good coffeeshops.

Research Agenda

My current work is organized around three connected questions:

Understanding and interpreting models

How can we better evaluate and understand the representations and behaviors that emerge in language models?

Representative work Mapping Language Models to Grounded Conceptual Spaces ↗

Improving alignment

How can we design learning objectives and feedback mechanisms that produce robust behavior when human values are diverse, conflicting, or underspecified?

Representative work Detecting and Controlling Sycophancy with Cascading Linear Features ↗

Anticipating emerging risks

How can models and multi-agent systems help us discover failure modes that static evaluations and human red teams have not yet anticipated?

Representative work League of LLMs: Creating Adversaries to Harden a Model ↗

Where I've Been

Google DeepMind, San Francisco
Safety, alignment, and reinforcement learning methods for LLMs.
2026–current
DeepMind, London
Language grounding, reinforcement learning, multi-agent and game-theoretic methods for LLMs.
2022–2025
Microsoft Research: Microsoft Turing Academic Program
Worked with Dean Carignan, Saurabh Tiwary, Pooya Moradi, Ali Alvi and others at MSR.
Summer 2021
DeepMind, London: Research Intern (Multi-agent Reinforcement Learning)
Worked with Angeliki Lazaridou, Edward Hughes, Ian Gemp and Yoram Bachrach.
Summer 2020
Google AI, Mountain View: Research Intern (Vision and Language Reinforcement Learning)
Worked with Alex Ku and Jason Baldridge.
Summer 2019
Johns Hopkins University: Jelinek Summer Workshop on Speech and Language Technology (JSALT)
Worked with Ellie Pavlick, Sam Bowman and Tal Linzen.
Summer 2018
Max Planck Institute: Cornell, Maryland, Max Planck Pre-doctoral Research School (CMMRS)
Summer 2018
University of Pennsylvania: Undergraduate Researcher
Worked with Ani Nenkova and Byron Wallace.
Summer 2017–18
Princeton University: Program in Algorithmic and Combinatorial Thinking (PACT)
Led by Rajiv Gandhi.
Summer 2016
Rutgers University: Undergraduate Researcher
Worked with Rajiv Gandhi.
2016–2018

Recent

September 2024
I'm giving a talk on safety and ethical implications of LLMs at the M2L Summer School .
July 2024
I'm giving a talk at a CogSci workshop on understanding the conceptual structure of language models.
June 2024
I'm at the Mexican NLP Summer School co-located with NAACL and on a panel talking about PhDs in NLP.
November 2023
I'm giving a guest lecture at Brown CS 1460 on grounded language learning.
October 2023
I'm at the DIMACS workshop talking about language models and multi-agent reinforcement learning can benefit from each other.
In today's garden path sentences: The prime number few.