Kimia Shaban

Toronto, Ontario, Canada

I am a PhD student in Computer Science at the University of Toronto, advised by Dr. Babak Taati. I work on computer vision and vision-language models. My PhD is funded by an NSERC Canada Graduate Scholarship (CGS-D).

I am a Faculty Affiliate Researcher at the Vector Institute and a Research Appointee at the KITE Research Institute at the University Health Network, where I work with the KITE Aging team on the Ambient project.

Research

My recent work asks:

Can diffusion transformers generate at resolutions they were never trained on, without retraining? SEGA scales attention per RoPE frequency band using the spectral energy of the latent at each denoising step. Accepted to NeurIPS 2026.

Before my PhD, I worked on machine learning for quantum field theory and combinatorics: predicting Feynman periods in φ⁴-theory with Dr. Paul-Hermann Balduf (JHEP 2024), the statistics of subdivergence-free Feynman integrals (arXiv), and witnesses in Covtree with Dr. Rowan Yeats (arXiv). The full list is on the publications page.

Beyond research

I run our group’s journal club, a series of paper discussions and guest talks that has been going since 2019. I set the schedule, invite speakers, and maintain the page. We meet on Tuesdays at the Vector Institute.
I co-organized the MoCha Workshop on Human Motion Challenges in Real-World and Clinical Settings at ECCV 2026.
I have been a teaching assistant for a dozen courses at Toronto and Waterloo, including Lead TA for CSC110 and CSC111, and received the C&O Department Outstanding TA Award at Waterloo. See teaching.

Background

I completed an MMath in Combinatorics & Optimization at the University of Waterloo, supervised by Dr. Rowan Yeats. My thesis is on combinatorial aspects of Feynman integrals and causal set theory. Before that I did a BMath at Waterloo, with a research term on deep Q-learning for antichains in Coxeter groups with Prof. Oliver Pechenik.

I grew up in Waterloo, Ontario. Outside of research I play piano, hike with my dog, and play Tetris.

news

Sep 24, 2026	SEGA was accepted to NeurIPS 2026!
Sep 09, 2026	Attended ECCV 2026 in Malmö, Sweden, where I co-organized the MoCha Workshop: Human Motion Challenges in Real-World and Clinical Settings.
May 28, 2026	New preprint: SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers — training-free high-resolution synthesis for diffusion transformers. Project page here.
Mar 02, 2026	New preprint: When Does RL Help Medical VLMs? — disentangling vision, SFT, and RL gains. Project page here.
Sep 01, 2025	Started my PhD in Computer Science at the University of Toronto, advised by Dr. Babak Taati. Also joined the Vector Institute as a Faculty Affiliate Researcher and KITE UHN as a Research Appointee.

selected publications

NeurIPS
SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers

Javad Rajabi, Kimia Shaban, Koorosh Roohi, and 2 more authors

In Advances in Neural Information Processing Systems (NeurIPS), 2026

Abs arXiv Bib HTML Website

Diffusion transformers (DiTs) have emerged as a dominant architecture for text-to-image generation, yet their performance drops when generating at resolutions beyond their training range. Existing training-free approaches mitigate this by modifying inference-time attention behavior, often through Rotary Position Embeddings (RoPE) extrapolation combined with attention scaling. However, these strategies apply a uniform and content-agnostic scaling across RoPE components with distinct frequency characteristics, inducing a trade-off between preserving global structure and recovering fine detail. We introduce SEGA, a training-free method that dynamically scales attention across RoPE components according to the latent’s spatial-frequency structure at each denoising step. This adaptive scaling improves both structural coherence and fine-detail fidelity. Experiments show that SEGA consistently improves high-resolution synthesis across multiple target resolutions, outperforming state-of-the-art training-free baselines.
@inproceedings{rajabi2026sega, title = {{SEGA}: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers}, author = {Rajabi, Javad and Shaban, Kimia and Roohi, Koorosh and Lindell, David B and Taati, Babak}, year = {2026}, booktitle = {Advances in Neural Information Processing Systems (NeurIPS)}, }
Preprint
When Does RL Help Medical VLMs? Disentangling Vision, SFT, and RL Gains

Ahmadreza Jeddi, Kimia Shaban, Negin Baghbanzadeh, and 4 more authors

arXiv preprint, 2026

Abs arXiv Bib HTML Website

We study when reinforcement learning provides meaningful gains over supervised fine-tuning for medical vision-language models, disentangling the contributions of vision encoders, SFT, and RL to model performance.
@article{shaban2026medbridgerl, title = {When Does RL Help Medical VLMs? Disentangling Vision, SFT, and RL Gains}, author = {Jeddi, Ahmadreza and Shaban, Kimia and Baghbanzadeh, Negin and Sharan, Natasha and Moturu, Abhishek and Dolatabadi, Elham and Taati, Babak}, year = {2026}, journal = {arXiv preprint}, }

JHEP

Predicting Feynman periods in φ^4-theory

Paul-Hermann Balduf and Kimia Shaban

Journal of High Energy Physics, Nov 2024

Abs DOI Bib HTML

@article{balduf2024feynman,
  title = {Predicting {F}eynman periods in $\phi^4$-theory},
  author = {Balduf, Paul-Hermann and Shaban, Kimia},
  journal = {Journal of High Energy Physics},
  year = {2024},
  month = nov,
  publisher = {Springer},
  doi = {10.1007/JHEP11(2024)038},
}