Now: Research intern at Microsoft Research Montréal
PhD candidate · Mila & Université de Montréal
I study how large language models generate text, act as agents and stay safe.
I'm a PhD candidate at Mila and Université de Montréal, advised by Sarath Chandar and Marc-Alexandre Côté, and currently a research intern at Microsoft Research Montréal. I work on large language models and the agents built on them, including safety methods that keep models helpful, test-time scaling for coding agents and diffusion language models, and multi-token prediction. Previously, I was a student researcher at ServiceNow Research and did my M.Sc. on offline reinforcement learning at the University of Alberta and Amii.
Latest first
I'm co-organizing REALM, the Second Workshop for Research on Agent Language Models, at EMNLP 2026 in Budapest on October 29.
Dialectics of Alignment was accepted at NeurIPS 2026.
The FrogNano technical report is out. I'm a core contributor and worked on test-time scaling for the agent.
New preprint showing that probabilistic calibration is a trainable capability in language models.
New preprint, REAM, on merging experts to compress mixture-of-experts LLMs.
Started a research internship at Microsoft Research Montréal, working on coding agents.
Joined ServiceNow Research as a student researcher.
Started my PhD at Mila and Université de Montréal with Sarath Chandar, supported by a DIRO Excellence Scholarship.
Current directions
Safety training that removes unsafe data or teaches refusals often leaves models overly cautious. I study how to keep that knowledge inside the model and control when it is used, and I'm now working on safety alignment for diffusion language models.
Small software-engineering agents that can run on modest hardware. I worked on test-time scaling for FrogNano, a 4B coding agent trained with reinforcement learning, and I'm working on context compaction for coding agents at Microsoft Research.
Language models that write many tokens at once, by denoising a whole sequence or by predicting several tokens per step. I study test-time scaling for diffusion language models and the noise structure of continuous diffusion LMs, and I'm benchmarking multi-token prediction methods and their scaling laws.
Making large models cheaper to run and more reliable. This includes compressing mixture-of-experts models by merging experts, and training models to sample from the distributions users ask for.
I've also worked on small language agents that learn sub-goals from an LLM (CoLLAs 2024 oral), on faster reinforcement learning by generalizing across state subspaces (TNNLS, Evolving Systems), and on how brains and neural networks represent language and value (EMNLP, eLife).
10 papers · Google Scholar
SafeMoE keeps unsafe knowledge in separate experts and learns, from as few as 100 examples, when to route to them, giving answers that are safer and more informative than standard alignment baselines.
A 4B-parameter coding agent trained only with reinforcement learning, on tasks synthesized online to match what the current model can learn.
Fine-tuning on synthetic sampling tasks teaches language models to follow requested randomness, and the skill carries over to unseen distributions and open-ended generation.
Merging experts gives better compressed mixture-of-experts LLMs than pruning them.
Distills an LLM's planning into a 770M-parameter hierarchical agent that needs no LLM calls at inference and beats imitation learning by 16.7 points on ScienceWorld.
fMRI and neurofeedback experiments show that value signals guide which abstract representations people use as they learn.
A character-level LSTM's representations still relate to human EEG responses for some kinds of nonsensical language.
Clusters the state space so an agent generalizes only from subspaces that fit its current state, which speeds up reinforcement learning.
Combines learning in feature subspaces with the full state space through confidence intervals, speeding up early model-based reinforcement learning.
2018 – now
Microsoft Research · Montréal
On the Froggy team, working on coding agents, including test-time scaling for FrogNano and context compaction for coding agents.
Mila & Université de Montréal
Advisors: Sarath Chandar and Marc-Alexandre Côté.
Mila – Quebec AI Institute · Montréal
Language-based agents.
Amii · Edmonton
Offline reinforcement learning.
University of Alberta & Amii · Edmonton
Advisors: Alona Fyshe and Martha White.
ATR · Kyoto, Japan
Modelling EEG data with reinforcement learning, with Aurelio Cortese.
Cold Spring Harbor Asia · Suzhou, China
Awards and community
When I'm not programming, I spend my time with my sweetheart, Saleh, whose way of understanding the world never stops amazing me. In summer I enjoy playing tennis, badminton and volleyball, and in winter I go skiing and ice skating. I also love art, and whenever I have time I paint with pastels.
I speak Farsi (native), English (full professional proficiency) and French (intermediate).
The best way to reach me is by email.
