Amir Eskandari

Portrait of Amir Eskandari
Plate I. The author.

I am a PhD candidate in the School of Computing at Queen’s University in Ontario, Canada, working on the personalization of large language models. I am supervised by Dr. Farhana Zulkernine and Dr. Jordan Poppenk. I am also a PhD trainee at Connected Minds CFREF.

Prior to my PhD, I was a graduate research assistant at AUT. I proudly hold an M.Sc. degree from Amirkabir University of Technology and a B.Sc. degree from IKIU, both in Electrical Engineering. During my master’s, I worked on multivariate time-series imputation using GNNs, supervised by Dr. Vahid Pourahmadi.

I love talking about science and technology. Shoot me an email if you’d like to discuss! You can also find me on Google Scholar, GitHub, LinkedIn and X.

I. News

DateEvent
Aug 2026I attended the RBC Borealis AI summer school, and our team won the challenge!
Jul 2026Our paper ElderBench, a benchmark of personalized open-source LLMs for older adults, was published at IEEE COMPSAC 2026 in Madrid.
Jul 2026Our system CASPAR for the MedExACT shared task on medical decision extraction was published at the BioNLP 2026 workshop at ACL.
Feb 2026Our paper ASMa on asymmetric spatio-temporal masking for skeleton action representation learning was published in Transactions on Machine Learning Research (TMLR).
Dec 2025My paper InfGraND, an influence-guided GNN-to-MLP knowledge distillation method, has been accepted for publication in Transactions on Machine Learning Research (TMLR).
Show 10 earlier entriesHide earlier entries
Sep 2025I started a new internship position as Machine Learning Associate at the Vector Institute!
Aug 2025Our paper CNN-CCA on anomaly detection in metro rail sensor data got published in Machine Learning with Applications (Elsevier)!
Jul 2025Our paper on self-supervised keypoint detection with distilled depth representations got accepted at ICMV 2025.
May 2025Two papers got accepted at IEEE COMPSAC 2025: SDA-GRIN for spatio-temporal time-series imputation, and a voice-adaptive LLM conversation bot.
Feb 2025I presented a poster at the Connected Minds annual retreat.
Aug 2024Submitted our survey on Transformer-based models to ACM Computing Surveys.
Aug 2024Our paper on GN2DI, a scalable GNN for spatial missing-data imputation in sensor networks, was accepted to IEEE FMLDS 2024.
May 2024I won the prestigious Connected Minds PhD Award!
Sep 2023I started my PhD in the School of Computing, Queen’s University!
Mar 2023I defended my master’s thesis on GNNs for multivariate time series with an excellent grade!

II. Research

My goal is one model that gives each person the answer that suits them. I am exploring different approaches to personalization, including retrieval, post-training (RL and SFT) and test-time scaling (Fig. 1). Personalization also comes with practical constraints, such as efficiency and local deployment on the user’s own device; I keep these in view and work toward methods that respect them. My research broadly spans graph machine learning and LLM post-training.

Where personalization can enter a language model User signals (stated preferences, interaction history, feedback and context) feed three stages of a language model that turns a query x into a response y for user u: the context, through retrieval; the weights, through post-training with SFT and RL; and decoding, through test-time scaling. All three sit within a budget of efficiency and local deployment. THE USER stated preferences interaction history feedback context retrieval post-training (SFT, RL) test-time scaling query x CONTEXT what it reads WEIGHTS what it has learned DECODING how it answers yu within a budget: efficiency, memory, and local deployment on the user’s own device Where personalization can enter a language model User signals (stated preferences, interaction history, feedback and context) feed three stages of a language model that turns a query x into a response y for user u: the context, through retrieval; the weights, through post-training with SFT and RL; and decoding, through test-time scaling. All three sit within a budget of efficiency and local deployment. THE USER stated preferences interaction history feedback context x CONTEXT via retrieval WEIGHTS via post-training (SFT, RL) DECODING via test-time scaling yu within a budget: efficiency, memory, and local deployment on the user’s own device
Fig. 1—Where personalization can enter a language model. What we know about the user can shape what the model reads, what it has learned, or how it decodes, and every route has to fit a practical budget.

Definition 1 (Personalization). Given a query x and what we know about a user u, a personalized model aims for the answer that this user prefers,

yu* = arg maxy ru(x, y),(1)

where ru is the user’s own reward, rather than one answer for everyone.

III. Selected Publications

  1. A. Eskandari, A. Anand, E. Rashno, and F. Zulkernine, “InfGraND: An Influence-Guided GNN-to-MLP Knowledge Distillation,” Transactions on Machine Learning Research (TMLR), 2026. [pdf] [project] [code] [blog]
  2. A. Eskandari, J. Tao, F. Zulkernine, M. Morningstar, J. Poppenk, and B. Herrmann, “ElderBench: Benchmarking Personalized Open-Source LLMs for Older Adults,” IEEE 50th Annual Computers, Software, and Applications Conference (COMPSAC), 2026.
  3. A. Anand, A. Eskandari, E. Rashno, and F. Zulkernine, “ASMa: Asymmetric Spatio-temporal Masking for Skeleton Action Representation Learning,” Transactions on Machine Learning Research (TMLR), 2026. [pdf]
  4. A. Eskandari, A. Anand, D. Sharma, and F. Zulkernine, “SDA-GRIN for Adaptive Spatial-Temporal Multivariate Time Series Imputation,” IEEE 49th Annual Computers, Software, and Applications Conference (COMPSAC), 2025. [pdf] [project] [code]
  5. E. Rashno, A. Eskandari, A. Anand, and F. Zulkernine, “Survey: Transformer-based Models in Multimodal Data Processing,” ACM Computing Surveys (under review), 2024. [preprint]

All papers are listed on the publications page.