Mingyi Hong

The paper introduces Straggler-Aware Group Control (SAGC), a dynamic group-size controller that optimizes synchronous on-policy RL training by adapting group size to minimize delays caused by slow rollouts (stragglers), thereby improving wall-clock efficiency and model performance.

This paper studies the distribution of reinforcement learning (RL) adaptation across transformer layers in large language models and finds that training a single layer can recover most of the gains obtained during full RL training.

This paper constructs strictly convex quadratic problems and initial points for which the long Barzilai--Borwein method does not converge root-superlinearly.