Salah Chikhi

Preprint, 2026

Last-Iterate Performance of Gradient Descent and Relaxed Proximal Point via s-Composability

We show that the joint potential of s-composability, introduced to compose optimized stepsize schedules, also certifies sharp last-iterate performance. The constant schedule is s-composable at every horizon, which identifies the unique constant stepsize minimizing the worst-case final gradient norm of smooth convex gradient descent, and of relaxed proximal point through the Moreau envelope. We also obtain the exact last-iterate constant of the silver schedule.

Under review at ICLR 2027

Taylor Representations for Model-Free RL in Networked MDPs

In networked MDPs, local-critic Taylor coefficients decay exponentially with the graph distance to the farthest agent involved, so discarding distant-agent coefficients yields scalable local Taylor representations. Building on them, we propose a model-free actor–critic algorithm with finite-sample guarantees for a linear LSTD critic. Across three control benchmarks, it matches or outperforms spectral baselines while scaling efficiently to large graphs.

With Abdelhaq Chaoui, Asuman Ozdaglar and Saurabh Amin

NeurIPS 2026

Adaptive Scheduling Pipeline for Multi-Instance Asynchronous Reinforcement Learning

ASP is a bounded-lag asynchronous architecture for RL post-training of large language models. A queue-aware scheduling policy jointly reduces padding, skewness and sorting bubbles, enabling efficient multi-instance rollout. ASP improves end-to-end throughput by up to 3.22× and reduces iteration time by up to 76%, with no visible convergence degradation.

With Luis Henrique Zuin Ruiz (equal contribution), Entong Li and Zeng Li

Preprint, 2025

An Analytical Framework for the Levine Hats Problem: New Strategies, Bounds and Generalizations

In the Levine hat problem, n ≥ 2 players must each identify a black hat on their own infinite stack, seeing only their teammates' stacks. Representing strategies as measurable functions, we construct a recursive strategy 𝒮₅ that attains the conjectured optimum 7/20 for two players, and we show that with uncountably infinite stacks the optimal winning probability is exactly 1/2.

With Clément Bouquet (equal contribution), Timothé Charles, Yanghao Zhou and Eric Wang