Alexia Jolicoeur-Martineau

Summary: Research Scientist at Samsung SAIL Montréal. Lead author of TRM (Tiny Recursive Model) and co-author of HRM (Hierarchical Reasoning Model). PhD from Mila/Université de Montréal. Research focus: recursive reasoning, deep supervision, ACT, small-model reasoning for ARC-AGI. Emerging star in efficient reasoning architectures.


Profile

Field Details
Current Role Research Scientist, Samsung SAIL Montréal
PhD Mila / Université de Montréal (2022-2024)
Thesis "Recurrent Neural Networks for Reasoning" (or similar)
Key Papers TRM (2025, lead), HRM (2025, co-author)
Research Focus Recursive reasoning, deep supervision, ACT, ARC-AGI, small models
Google Scholar Profile
Twitter/X @ajolicoeurm
Website alexiajm.github.io

Significance for You

Perfect research match for your interests:

Your Interest Her Work
Recursive / latent space reasoning HRM/TRM: recursive latent updates
Adaptive computation / ACT Q-learning halting in HRM/TRM
Deep supervision Primary innovation in TRM (19%→39%)
Small efficient models 7M params beating 7B LLMs on ARC
World models / planning Recursive latent = implicit planning

PhD advisor potential:

  • Industry lab (Samsung) — may co-advise with academic partner (Mila/UdeM)
  • Early career → likely taking students
  • Publishes at top venues (ICLR, NeurIPS)
  • Montréal = can visit Mila ecosystem easily

Target tier: Tier 1 — exact research match, emerging leader, accessible lab.


Key Publications

Paper Year Role Innovation Your Wiki
TRM: Tiny Recursive Model 2025 Lead 7M params, flat recursion, deep supervision primary driver, 45% ARC-AGI-1 paper-trm
HRM: Hierarchical Reasoning Model 2025 Co-author H/L hierarchy, deep supervision, ACT, 40.3% ARC-AGI-1 paper-hrm
PhD work 2022-24 Lead Recurrent reasoning foundations

Research Themes (from papers)

  1. Deep supervision > hierarchy — TRM ablation: deep supervision gives +20% absolute; hierarchy only +3.3%
  2. ACT without replay buffers — Q-learning stability via Post-Norm + AdamW (bounded params)
  3. Inference-time compute scaling — Train M_max=8, test M_max=16 zero-shot gains
  4. Flat recursion beats hierarchy — TRM (single net) > HRM (two nets) with 4× fewer params
  5. Small models for reasoning — 7M params competitive with LLMs on ARC with 1000× less compute

Contact Strategy

When: Summer 2025 (pre-PhD applications) How: Email + attach your best paper/work (DSA-ViT project page, any preprints) Angle:

"Your TRM paper showed deep supervision is the key driver for recursive reasoning. I've been adapting DSA (DeepSeek Sparse Attention) to Vision Transformers for efficient long-range reasoning. I'd love to explore: (1) TRM-style deep supervision for visual reasoning / video understanding, (2) ACT for adaptive visual computation, (3) recursive latent reasoning for world models."

Follow-up: If responsive, propose virtual meeting; mention Montréal visit possibility.


Related Wiki Pages


Sources

  • arxiv-2510.04871: TRM paper (lead author)
  • arxiv-2506.21734: HRM paper (co-author)