Luke Marris

Timeline

1. 2018–2019

   Capture the Flag agents

   He is one of 18 authors of a 2018 paper showing that populations of reinforcement-learning agents, trained concurrently from thousands of parallel matches, reach human-level play in Quake III Arena Capture the Flag from pixels and game points. The paper appears in Science in 2019 as Human-level performance in 3D multiplayer games with population-based reinforcement learning.

2. 2020

   Backpropagation and the brain

   In 2020, he co-authors a Nature Reviews Neuroscience perspective with Timothy Lillicrap, Adam Santoro, Colin Akerman and Geoffrey Hinton. The article argues that feedback connections may induce neural activities whose differences locally approximate the error signals of backpropagation. His affiliation is DeepMind in London.

3. 2021

   Correlated equilibrium meta-solvers

   As first author at ICML 2021, he proposes Joint Policy-Space Response Oracles (JPSRO), an algorithm for training agents in n-player, general-sum extensive-form games that provably converges to an equilibrium. The paper, written with Paul Muller, Marc Lanctot, Karl Tuyls and Thore Graepel, also introduces the Maximum Gini Correlated Equilibrium for selecting among correlated equilibria.

4. 2022

   The Neural Equilibrium Solver

   His first-author paper Turbocharging Solution Concepts, written with Ian Gemp, Thomas Anthony, Andrea Tacchetti, Siqi Liu and Karl Tuyls, introduces the Neural Equilibrium Solver, an equivariant neural network that approximately solves the space of all normal-form games of a fixed shape. It appears at NeurIPS 2022. He is also among the authors of NeuPL, a method for neural population learning, at ICLR 2022.

5. 2024

   An ICLR honorable mention

   Approximating Nash Equilibria in Normal-Form Games via Stochastic Optimization, written with Ian Gemp and Georgios Piliouras, proposes the first loss function for approximate Nash equilibria of normal-form games that allows unbiased Monte Carlo estimation. The paper receives an Honorable Mention and is presented as an oral at ICLR 2024.

6. By 2024

   Cambridge degrees and a UCL PhD

   His University College London PhD thesis, Multiagent Training in N-Player General-Sum Games, appears in 2024. It builds fast neural network solvers for equilibria in normal-form games and population-based learning algorithms for training policies in n-player general-sum extensive-form games. He holds first-class bachelor's and master's degrees in Information Engineering from the University of Cambridge.

7. 2025

   Gemini 2.5 and deviation ratings

   In 2025, he is an author of the Gemini 2.5 report, which introduces the Gemini 2.X model family of Gemini 2.5 Pro and Gemini 2.5 Flash alongside the earlier Gemini 2.0 Flash and Flash-Lite. As first author of Deviation Ratings, written with Siqi Liu, Ian Gemp, Georgios Piliouras and Marc Lanctot, he introduces the first N-player general-sum clone invariant rating, based on coarse correlated equilibria and explored on LLM evaluation.

8. As of October 2026

   Staff research engineer at Google DeepMind

   He is a Staff Research Engineer at Google DeepMind in London. His research interest is training deep reinforcement-learning agents, at scale, in many-player mixed-motive games with algorithms that select and compute equilibria. A 2026 arXiv paper, Deep Incentive Design with Differentiable Equilibrium Blocks, written with V. Thoma and Georgios Piliouras, proposes a differentiable framework for designing multi-agent interactions with desirable equilibrium outcomes.