Ailar Mahdizadeh

I work on multimodal and vision–language models, long-context video understanding, foundation-model post-training, and efficient inference.

ailar.mahdizadeh@ubc.ca GitHub LinkedIn

About me

I'm a PhD candidate in the Electrical and Computer Engineering department at the University of British Columbia and a graduate researcher at the Vector Institute for AI, working with Leonid Sigal in the UBC Vision group. My research spans multimodal large language models, streaming and long-context video understanding, foundation-model post-training, knowledge distillation, KV-cache compression, and robustness under distribution shift. I have interned at Global Relay and Roche.

Before my PhD, I completed an MASc in 2024 in Electrical and Computer Engineering at the University of British Columbia in the TEA lab with Xiaoxiao Li.

Beyond research, I collect Lego, make journals, and sing Happy Birthday to the occasional bear on BC trails. Five star reviews from the bears so far.

Research highlights

2× lower VRAM at comparable accuracy on long video

CoRDS

Coreset-based Representative and Diverse Selection for streaming video understanding. One principled, training-free selector picks visual tokens that are both representative of the value-norm-weighted importance landscape and diverse across the residual subspace, compressing the KV cache block-by-block so long videos never materialize a full cache. Plugs into Qwen2-VL and Qwen2.5-VL with no per-benchmark tuning.

arXiv ↗Code ↗

VLMsKV-cache compressionlong video

Honors & awards

2024MUSIC Graduate Student Program awardee, UBC

2024Vector Institute for AI Research Scholarship

2023UBC Graduate Support Initiative award for academic excellence