Rishabh Agarwal

Rishabh Agarwal is a reinforcement learning researcher whose work includes statistical tools for evaluating deep RL agents and on-policy distillation for language models. A Computer Science and Engineering graduate of IIT Bombay, he spent a year with Geoffrey Hinton's team at Google Brain in Toronto, and in 2021 he won a NeurIPS Outstanding Paper Award with his Mila PhD advisors Aaron Courville and Marc Bellemare, among others. In 2023, he posted a paper on on-policy distillation of language models that introduced Generalized Knowledge Distillation. Following a stint as a staff research scientist at Google DeepMind, he joined Meta in April 2025 and left in August 2025. As of October 2026, he is a founding member at Periodic Labs and an adjunct professor at McGill University.

Now at: Periodic Labs and McGill University
Badges: Founder
Bridge

Timeline

2019–2020

Offline reinforcement learning with REM

A Computer Science and Engineering graduate of IIT Bombay, he posts "An Optimistic Perspective on Offline Reinforcement Learning" with Dale Schuurmans and Mohammad Norouzi in 2019, and it appears in the ICML 2020 proceedings. The paper presents Random Ensemble Mixture (REM), a Q-learning algorithm, and shows that agents trained offline on a DQN replay dataset from 60 Atari games outperform the fully trained DQN agent. He also spends a year with Geoffrey Hinton's team at Google Brain in Toronto.
2021

A NeurIPS Outstanding Paper

In 2021, he wins a NeurIPS Outstanding Paper Award with Max Schwarzer, Pablo Samuel Castro and his Mila PhD advisors Aaron Courville and Marc Bellemare for "Deep Reinforcement Learning at the Edge of the Statistical Precipice". The paper recommends stratified bootstrap confidence intervals, performance profiles and interquartile means for comparing algorithms, and comes with the open-source library rliable.
2023

On-policy distillation for language models

In 2023, he posts "On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes" with six coauthors, including Nino Vieillard and Olivier Bachem. The paper introduces Generalized Knowledge Distillation, which trains a student model on its own generated output sequences using feedback from the teacher.
April 2025

From DeepMind to Meta

In April 2025, he leaves Google DeepMind, where he has been a staff research scientist and a core contributor to the Gemma and Gemini models, and joins Meta. He moves to Meta's superintelligence team and works on reinforcement learning and reasoning.
August 2025

Leaving Meta after five months

In August 2025, he announces his departure from Meta after about five months. In his farewell post, he praises the talent and compute density of the lab and writes that he felt the pull to take on a different kind of risk.
As of October 2026

Founding member at Periodic Labs

He is a founding member at Periodic Labs, an AI research and deployment company creating models and autonomous labs to accelerate science, and an adjunct professor at McGill University.