Skander Moalla

My research focuses on scalable learning from experience, driven by three core ingredients: plasticity for continual learning, off-policy/offline reinforcement learning for efficiency, and diversity (parallel) and agentic systems (sequential) for exploration and test-time scaling.

My work includes co-leading the post-training of an open-source 70B LLM (Apertus, ACL’26), leveraging diversity to improve test-time scaling and exploration in RL (Google DeepMind Intern, Gemma post-training), developing an agentic harness for scaling AI theorem proving in Lean (h1, Meta FAIR), deriving an offline RL algorithm for LLMs (QRPO, NeurIPS’25), and exposing the connection between plasticity, trust-region, and off-policy collapse (No Rep, No Trust, NeurIPS’24).

I have developed several codebases from scratch with significant infrastructure work such as a scalable code execution sandbox for QRPO, a scalable post-training pipeline for Apertus 70B, and a reproducible ML template. I am also proud of conducting award-winning reproducibility studies (ReScience 2023).

I am considering full-time positions for winter 2026/2027. I bring a strong background in both research and engineering.

Background

I was a Research Scientist Intern at Meta FAIR (Paris) hosted by Julia Kempe and Rémi Munos in the Foundations of Reasoning / Math Superintelligence Team, working on agentic in-context adaptation, self-improvement, and asymmetric self-play in formal math (Lean). I was also a PhD student researcher at Google DeepMind (Paris) with the Gemma post-training team hosted by Alexandre Ramé, working on diversity to improve test-time scaling and sampling in reinforcement learning. I also interned as an applied scientist at Quincus to optimize middle-mile logistics using RL, and as a software engineer at Amazon building AI tools for Amazon Transportation Services.

I hold an MSc in Advanced Computer Science from the University of Oxford (Reuben College inaugural cohort 2021-2022). My MSc thesis on Multi-Agent RL and policy gradient methods was supervised by Mingfei Sun and Shimon Whiteson at the Whiteson Research Lab (WhiRL).

I completed a BSc in Mathematics and Computer Science from École Polytechnique (inaugural cohort 2017-2020) and was a visiting student at EPFL, the University of Toronto, and Stanford University, where I focused on software engineering and entrepreneurship.

At EPFL, I was the president of CUBAliente (2024), the Latin social dancing association. At Oxford, I led the Careers team of the Oxford Artificial Intelligence Society (OxAI) (2022) and played as a libero for the men’s Blues volleyball team (2021-2022). At École Polytechnique, I was elected to the first board of the L’ORE (2018), the Bachelor’s program student society.