Thomas Jiralerspong

I am a PhD student in computer science co-supervised by Yoshua Bengio and Guillaume Lajoie at Mila and Université de Montréal.

I am also currently an AI research fellow at Anthropic, working on mechanistic interpretability of emergently misaligned models.

My research has been funded by a Vanier Canada Graduate Scholarship, NSERC Canada Graduate Scholarship, and FRQNT Scholarship.

I have a wide array of interests, most of which I have been fortunate enough to work on:

• LLMs/VLMs:
  • LLM agents (internship at Occam AI)
  • LLM internals and mechanistic interpretability (paper on compositionality in LLMs)
  • LLM applications (paper on causal graph discovery with LLMs)
• Neuroscience/Cognitive science/Psychology inspired AI:
  • Input-driven learning (paper on bias-only learning)
• Modularity:
  • Compositionality (papers on complexity-based theory of compositionality and compositionality in LLMs)
  • Discrete Representations (paper on discovering discrete subgoals for RL)
• Model-based AI:
  • Causality (papers on causal graph discovery with LLMs and causal imputation)
  • Model-based reinforcement learning (paper on temporally extended tree-search planning)
• AI for good:
  • AI for Healthcare (paper on RL for Mechanical Ventilation)
  • AI for Climate Change (paper on RL for HVAC Control)
  • AI for Autonomous Driving (Internship at Waabi)
  • AI for Drug Discovery

In my free time, I enjoy traveling, watching/analyzing good movies, and writing sad songs (check out my (very) amateur music here)!

Reach out at thomas.jiralerspong@mila.quebec if there is anything you want to discuss, I’m always happy to talk!