Hamed Shirzad

Hamed Shirzad

Computer Science PhD candidate, UBC  ·  Student Researcher, Google Research

Google Scholar GitHub LinkedIn Twitter CV (PDF)

I'm a Computer Science PhD candidate at the University of British Columbia, advised by Prof. Danica Sutherland, and a Student Researcher at Google Research. My work lives at the meeting point of learning algorithms and architectures that actually scale.

Right now I'm focused on post-training large language models with reinforcement learning, getting models to improve themselves through self-play rather than leaning only on human labels. Earlier, a big chunk of my research was about making transformers on graphs scale to real problem sizes, and building more trustworthy ways to evaluate and generate structured data. I like methods that are motivated by theory but earn their keep empirically, and I'm increasingly drawn to pointing them at scientific questions like drug discovery.

Research Interests

LLM post-training Reinforcement learning Self-play & self-improvement Efficient transformers Graph machine learning Generative models Representation learning & evaluation ML for drug discovery

News

May 2026

SeedER, our RL approach to knowledge-graph retrieval, is out.

2026

TxPert is published in Nature Biotechnology. It predicts transcriptomic responses to unseen genetic perturbations.

Feb 2026

Started as a Student Researcher at Google Research, working on reinforcement-learning post-training for Gemma models.

Aug 2025

Gave a tutorial at MLSP 2025 on geometry in graph learning.

May 2025

Gave a keynote at GSP 2025 on scaling graph transformers with sparse and sparsified attention.

Sep 2024

Even Sparser Graph Transformers accepted at NeurIPS 2024.

Research Experience

Student Researcher · Google Research
Feb 2026 – Present · Remote
RL post-training of Gemma models via self-play and self-improvement.
  • Developing reinforcement-learning post-training methods in which language models improve through self-play, generating and learning from their own training signal.
  • Fine-tuning Gemma 4 models up to 31B parameters with RL fine-tuning techniques, implemented in JAX, Flax, and the Tunix post-training library.
Research Intern · Valence Labs (Recursion), Virtual Cell
Mar 2025 – Dec 2025 · Montreal, Canada
Knowledge-graph models for biology; contributed to TxPert and led SeedER.
  • Co-developed knowledge-graph-based structured-prediction models of transcriptomic perturbation effects, contributing to TxPert (Nature Biotechnology, 2026).
  • Developed SeedER, an RL-based seed-and-expand method for retrieval from biological knowledge graphs (first-author paper under review at NeurIPS 2026).
  • Worked in an interdisciplinary team connecting machine-learning methods with large-scale biological prediction for drug discovery.
Research Intern · Borealis AI (now RBC Borealis)
Sep 2023 – Feb 2024 · Vancouver, Canada
Diffusion models as self-supervised backbones for irregular time series.
  • Developed conditional diffusion models as self-supervised learning backbones for irregular event-sequence and time-series data (first-author paper at the Time Series for Health workshop, ICLR 2024).
Research Intern · Autodesk AI Lab
Apr 2021 – Sep 2021 · Remote
Contrastively-learned evaluation metrics for graph generative models.
  • Developed contrastively-learned evaluation metrics for graph generative models, resulting in a first-author NeurIPS 2022 paper.
Research Intern · Borealis AI
Apr 2020 – Mar 2021 · Remote
Permutation-invariant graph generation via tree decomposition.
  • Developed permutation-invariant graph generation via tree decomposition with Prof. Greg Mori and Dr. Hossein Hajimirsadeghi, leading to the first-author AISTATS 2022 paper TD-Gen.
Research Intern · Max Planck Institute for Intelligent Systems
Jul 2018 – Sep 2018 · Tübingen, Germany
Kernelized Stein discrepancy and its applications.
  • Worked on kernelized Stein discrepancy and its applications under Prof. Krikamol Muandet.

Selected Publications

∗ equal contribution. Full list on Google Scholar.

SeedER: Seed-and-Expand Retrieval from Knowledge GraphsUnder review · NeurIPS 2026

H. Shirzad, F. Wenkel, D. Beaini, D. J. Sutherland, E. Noutahi

TL;DRMost knowledge-graph retrieval grabs a fixed blob of neighbors around your query and hopes the answer is in there. We instead train a GNN with RL to start from a few seed nodes and decide, step by step, which directions are actually worth expanding into, so it reaches the relevant facts without dragging along all the noise.
arXiv

TxPert: Using Multiple Knowledge Graphs for Prediction of Transcriptomic Perturbation EffectsNature Biotechnology 2026

F. Wenkel, W. Tu, C. Masschelein, H. Shirzad, et al.

TL;DRIf you knock out a gene a cell has never had knocked out before, how does its gene expression shift? We lean on several biological knowledge graphs at once so the model can reason about relationships it hasn't seen directly. That's what lets it generalize to genuinely unseen perturbations instead of just memorizing the training set.
Paper Code Blog

Even Sparser Graph TransformersNeurIPS 2024

H. Shirzad, H. Lin, B. Venkatachalam, A. Velingker, D. Woodruff, D. J. Sutherland

TL;DRGraph transformers get expensive fast because attention wants to connect everything to everything. We show you can be far more aggressive about sparsifying it than people assumed: train a tiny throwaway model, watch which connections it actually uses, keep only those. The full-size model trained on that sparse graph still matches accuracy while scaling to much larger graphs.
arXiv Code

Exphormer: Sparse Transformers for GraphsICML 2023

H. Shirzad∗, A. Velingker∗, B. Venkatachalam∗, D. J. Sutherland, A. K. Sinop

TL;DRFull attention on a graph costs quadratically and falls over on big graphs. We build attention out of expander graphs plus a few virtual nodes, which keeps information flowing globally across the whole graph but at linear cost. It turned into a widely-used backbone for graph transformers and got written up on the Google Research blog.
arXiv Code Google blog

Evaluating Graph Generative Models with Contrastively Learned FeaturesNeurIPS 2022

H. Shirzad, K. Hassani, D. J. Sutherland

TL;DRHow do you actually tell whether a graph generator is any good? The usual hand-picked statistics quietly miss a lot. We show that features from a contrastively-trained GNN make a much more reliable yardstick, and that combining them with classic descriptors beats either alone.
arXiv Code

TD-Gen: Graph Generation Using Tree DecompositionAISTATS 2022

H. Shirzad, H. Hajimirsadeghi, A. H. Abdi, G. Mori

TL;DRWe generate a graph by first building a tree of node-clusters and then filling in edges cluster by cluster. Because the tree backbone is permutation-invariant, this sidesteps the node-ordering headaches that trip up most sequential graph generators. We also argue for evaluating such models by likelihood rather than the shaky statistics people usually reach for.
arXiv

A Theory for Compressibility of Graph Transformers for Transductive LearningMLComp Workshop · NeurIPS 2024

H. Shirzad, H. Lin, B. Venkatachalam, A. Velingker, D. Woodruff, D. J. Sutherland

TL;DRThe theory companion to our sparsification work: we prove why graph transformers for transductive tasks can be compressed so aggressively without losing much, turning an empirical trick into something with guarantees behind it.
arXiv

Conditional Diffusion Models as a Self-supervised Backbone for Irregular Time SeriesTS4H Workshop · ICLR 2024

H. Shirzad, R. Deng, H. Zhao, F. Tung

TL;DRTraining a conditional diffusion model turns out to rhyme with masked auto-encoding. We use that to learn representations of messy, irregularly-sampled time series (think health records), and the model's hidden states transfer surprisingly well to downstream prediction tasks.
Paper

Invited Talks

May 2025

KeynoteScaling Graph Transformers with Sparse and Sparsified Attention at GSP 2025

Aug 2025

TutorialGeometric and Topological Representation Learning at MLSP 2025 · slides & materials

Contact

The fastest way to reach me is email: . I'm always happy to talk about post-training, efficient architectures, or ML for the sciences.

I'm also on the lookout for research scientist roles where I can keep pushing on these questions. If that overlaps with what your team is working on, I'd love to hear from you.

Google Scholar GitHub LinkedIn CV (PDF)