Samin Yeasar Arnob

Samin Yeasar Arnob

I am working at Cohere on improving LLM reasoning (300B+) models.

I am a (soon to graduate) PhD student from McGill University and Mila, advised by Professor Doina Precup. My PhD focuses on efficient and adaptive learning in large neural networks by discovering and exploiting sparse subspaces inspired by brain-like neural pathways. I study how overparameterized models in reinforcement learning and large language models can learn effectively using only a small subset of parameters, improving energy efficiency, scalability, and continual learning. My work introduces Neural Pathways for multitask RL, Sparse-Reg to improve sample efficiency and robustness in offline RL, and Sparse Adapters for modular and compositional LLM fine-tuning and model merging.

Email / CV / Google Scholar / X / Github

Research Affiliations


Cohere Intern of Technical Stuff 09/2025 – Present	McGill University PhD student 01/2020 – Present	Mila Student Researcher 01/2020 – Present

Microsoft Research, MTL Student Researcher 03/2024 – 04/2025	Microsoft Research, NYC Research Intern 04/2023 – 08/2023	Ubisoft Research Intern 09/2021 – 08/2022

Research and Work Experience

Cohere

Fall 2025, Winter 2026

Intern of Technical Stuff

Building controlled-synthetic reasoning pipelines and tooling for large-scale LLM reasoning and structured training recipes.

LLMs Reasoning Data pipeline RLVR / post-training

Microsoft Research, Montréal

Summer 2024, Fall 2024, Winter 2025

Student Researcher

Developed Sparse Adapters for modular, parameter-efficient fine-tuning LLMs and scalable model merging.

LLMs Sparse adapters Model merging PEFT

Microsoft Research, NYC

Summer 2023

Applied Research Intern

Built agentic exploration systems using hierarchical world modelling and latent world modeling for efficient large-scale system interaction.

Agentic system World models Hierarchical planning Exploration

Ubisoft

Fall 2021, Winter 2022, Summer 2022

Research Intern

conducted research on large map navigation for bots using imitation learning and offline RL, experimented with the GPT-2 model to enhance performance and generalization in automating bot navigation for future games.

Reinforcement Learning (RL) GPT2 Imitation learning Offline RL

Mila – Quebec AI Institute

2020 – Present

Student Researcher

Research with Doina Precup on sparse subspace optimization across reinforcement learning and large language models.

Reinforcement Learning (RL) LLMs Subspace learning Sparse learning

Research Interest

Reinforcement Learning: Imitation Learning, Offline Reinforcement Learning, Multitask Learning, Representation Learning, Hierarchical World Model
Large Language Model: Reasoning, RLVR, RLHF, Mixture of Experts (MoE), Improving mergeability of experts, Parameter-efficient finetuning (PEFT)

	Exploring Sparse Adapters for Scalable Merging of Parameter Efficient Experts Samin Yeasar Arnob, Zhan Su, Minseon Kim, Oleksiy Ostapenko, Esra Saleh, Riyasat Ohib, Doina Precup, Lucas Page-Caccia, Alessandro Sordoni COLM 2025, Project, Code Keywords: Sparse adapter, Parameter-efficient finetuning, Model merging, LLM Summary: We propose a simple, yet efficient sparse adapters training method as a building block for modular, parameter-efficient LLM finetuning, improving merging performance at scale.
	Efficient Reinforcement Learning by Discovering Neural Pathways Samin Yeasar Arnob, Riyasat Ohib, Sergey Plis, Amy Zhang, Alessandro Sordoni, Doina Precup NeurIPS 2024, Project, Code Keywords: Neural Pathways, Parameter-efficient training, (Online/Offline) RL, Multitask RL Summary: Learn specialized sparse subspaces for each RL agent that can co-exist, using ~5% of total parameters.
	Sparse-Reg: Improving Sample Complexity of Offline Reinforcement Learning using Sparse Regularization Samin Yeasar Arnob, Scott Fujimoto, Doina Precup RLDM 2025, Code Keywords: Offline RL, Sparsity, Regularization, Sample Complexity, Continuous Control Summary: Sparse regularization to mitigate overfitting in small offline datasets and improve continuous control performance.
	Importance of Empirical Sample Complexity Analysis for Offline Reinforcement Learning Samin Yeasar Arnob, Riashat Islam, Doina Precup NeurIPS 2021 (Offline RL Workshop), Code Keywords: Offline RL, Sample Complexity Summary: Study sample complexity as a robustness signal; highlight that many offline RL methods fail in low-data regimes.
	Single-Shot Pruning for Offline Reinforcement Learning Samin Yeasar Arnob, Riyasat Ohib, Sergey Plis, Doina Precup NeurIPS 2021 (Offline RL Workshop) Keywords: Offline RL, Sparse Networks, Pruning, Single-shot pruning Summary: First work to demonstrate that single-shot pruning can be effective for offline reinforcement learning.
	OAIRL: Off-policy Adversarial Inverse Reinforcement Learning Samin Yeasar Arnob ICML 2020 (Lifelong Learning Workshop), Code Keywords: RL, Transfer Learning, Inverse Reinforcement Learning Summary: Improve imitation performance and transfer knowledge under dynamic task changes.

The template is taken from here. Thanks so much!