July 14, 2025
Microsoft Research Vancouver &
Canadian Academic Community Workshop
08:00 – 14:00 PT
Location: Vancouver, B.C.
July 14, 2025
| Time | Agenda | Speakers |
|---|---|---|
| 08:00 – 09:00 | Check-in and Breakfast | |
| 09:00 – 09:05 | Opening remarks | Lidong Zhou, Corporate Vice President of Microsoft, Managing Director of Microsoft Research Asia |
| 09:05 – 09:10 | MSR Vancouver Lab Introduction | Peng Cheng, Microsoft Research Vancouver |
| 09:10 – 10:10 | Lightning session 1: Invited professors | Lightning Talk A: Automated vulnerability discovery and repair with LLMs Steve Ko, Simon Fraser University This is an umbrella project and we’re working on two things now. One is to design a vulnerability benchmark for Rust that is context- and contamination-aware. Another is to design a targeted concolic execution engine with LLM-based target selection. Lightning Talk B: Towards Efficient Long-Context LLMs Ke Li, Simon Fraser University LLMs need to support large context windows to understand complex prompts and to learn from complex examples at test time. However, a large context window yields many keys, which degrades the computational and memory efficiency of LLMs. Our work introduces new algorithms to address both challenges, achieving a speedup of 2.7-7.6x and a memory reduction of 2x with almost no loss in accuracy. Lightning Talk C: Certifiably Trustworthy Deep Learning in the era of Large Language Models Linyi Li, Simon Fraser University In this lightning talk, I am going to introduce our lab’s past and ongoing efforts on diagnosing and demystifying LLM’s learning behaviors towards improving its trustworthiness in a certifiable way. Topics may be extracting model’s sensitive features, attacking GUI agents, attribution guided data selection, and smoothness-bounded architectural design. |
| 10:10 – 10:20 | Coffee break and mingle | |
| 10:20 – 11:20 | Lightning session 2: Invited professors | Lightning Talk D: Cross-Lingual Knowledge Transfer with LLMs Vered Shwartz, University of British Columbia & Vector Institute Factual knowledge is often available online—for example, in Wikipedia—only in certain languages, depending on local relevance. Can multilingual LLMs like GPT-4 help present users with relevant facts from various language origins? Lightning Talk E: Is scaling current agent architectures the most effective way to build generalist agents? Souradeep Dutta, University of British Columbia This talk introduces REGENT, a semi-parametric agent that adapts to new environments using retrieval-augmented, in-context learning instead of large-scale fine tuning. By leveraging a retrieval-based bias, REGENT achieves strong generalization in robotics and games with far fewer parameters and training data. This approach outperforms current state-of-the-art generalist agents in both real and simulation experiments. Lightning Talk F: Building and Deploying LLM Agentic Systems with Compact and Efficient Models Xiaoxiao Li, University of British Columbia & Vector Institute Description Large language models (LLMs) have transformed AI, enabling breakthroughs across diverse applications—but their immense size severely restricts practical deployment. How can we preserve the extraordinary capabilities of LLMs while making them dramatically smaller and faster? In this talk, I will highlight our recent work on LLM pruning and self-improvement on small LLMs that not only achieve remarkable reductions in computational cost and memory usage but also create opportunities for LLMs to adapt seamlessly to diverse deployment scenarios. |
| 11:20 – 11:30 | Coffee break and group photo | |
| 11:30 – 12:30 | Panel discussion | This panel brings together voices from Microsoft Research and the Canadian academic community to explore the evolving landscape of AI research. The discussion will delve into the motivations that have shaped researchers’ journeys, the challenges they currently face, and the collaborative opportunities that lie ahead. Panelists: Ke Li, Simon Fraser University Linyi Li, Simon Fraser University Steve Ko, Simon Fraser University Souradeep Dutta, University of British Columbia Vered Shwartz, University of British Columbia & Vector Institute Xiaoxiao Li, University of British Columbia & Vector Institute Peng Cheng, Microsoft Research Vancouver Moderator: Soheil Abbasloo, Microsoft Research Vancouver |
| 12:30 – 12:35 | Closing remarks | Lidong Zhou, Corporate Vice President of Microsoft, Managing Director of Microsoft Research Asia |
| 12:35 – 14:30 | Lunch and networking |