Michael Noukhovitch

Michael Noukhovitch is an AI researcher whose work spans communication between agents and reinforcement learning for language models. He earned a software engineering degree at Waterloo in 2017 and a master's at Mila in 2019, then coauthored research on competitive communication, language drift, and model alignment. He proposed asynchronous RLHF to train on older samples while generating new ones, and he helped build RL infrastructure for Olmo 3 at the Allen Institute for AI. In 2026, he coauthored work on reward hacking and Never Give Up, an adaptive method that gives harder problems more compute. As of October 2026, he is a PhD candidate at Mila, associated with Université de Montréal and supervised by Aaron Courville.

Now at: Université de Montréal and Allen Institute for AI (Ai2)
Works: 17
Citations: 149
h-index: 5

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2017–2019

Waterloo and Mila degrees

In 2017, he earns a software engineering degree at the University of Waterloo and joins Team Frustra for Stray Photons, a Software Engineering capstone that uses global illumination as a game mechanic. He completes a master's at Mila in 2019.
2021

Communication under competition

He coauthors Emergent Communication under Competition in 2021. Its modified sender-receiver game shows that communication can emerge among agents in partially competitive settings.
2022

Language drift under competition

He coauthors Competition exacerbates Language Drift for the Machine Learning and the Evolution of Language workshop at JCoLE in 2022.
2023

Elastic Reset for alignment

In 2023, he coauthors Language Model Alignment with Elastic Reset. The method periodically resets an online model to its exponential moving average and resets the average to the initial model, reaching higher reward with less drift.
2024–2025

Asynchronous RLHF for language models

He proposes Asynchronous RLHF in 2024, separating generation from learning so new samples can be generated while older samples train the model. At ICLR 2025, the paper reports training LLaMA 3.1 8B about 40% faster than a synchronous run at matching performance.
2025

Building Olmo 3 at Ai2

At the Allen Institute for AI, he works as a research intern on Olmo 3, helps build its RL training infrastructure, and leads the RL-Zero effort. In November 2025, Olmo 3 is released as a fully open 7B and 32B language-model family whose model flow includes every stage, checkpoint, data point, and dependency.
2026

Gradient regularization for RLHF

In 2026, he coauthors Gradient Regularization Prevents Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards, an ICML paper that reports better results than a KL penalty across language-model RL experiments.
As of October 2026

PhD research at Mila

As of October 2026, he is a PhD candidate in artificial intelligence at Mila, associated with Université de Montréal and supervised by Aaron Courville. In September 2026, he coauthors Learning to Solve Hard Problems in RL for LLMs by Never Giving Up, which introduces adaptive sampling that uses fewer samples on easy problems and allocates more compute to harder problems.