Mohammad Sami Nur Islam

Hi! I am an MSc student in Computer Science at McGill/Mila, advised by Professor Doina Precup. Previously, I completed my undergrad at McGill.

Research

My research primarily focuses on reinforcement learning (RL), centered around scaling, plasticity, and representation dynamics. Recently, my work investigates offline-to-online RL in the absence of offline data during the online phase, using expressive frozen actor priors to prevent representation collapse and policy forgetting. Alongside this, I am exploring with Professor Benjamin Eysenbach how to scale unsupervised, goal-conditioned RL to complex control tasks without engineering dense reward functions or sub-tasks.

Previously, I have worked on designing reward functions for sparse reward environments using VLMs and improving exploration in deep continuous control with Thompson Sampling.

News