Larry, Yinxi Li
yinxi.li[at]uwaterloo.ca | SWAG Lab DC 2555
Amor fati, carpe diem
🦄 About Me | Currently Master’s Student in CS
Hey there, I’m Larry! Thanks for visiting my humble page. I’m currently a Computer Science M.Math. student at the University of Waterloo, advised by Pengyu Nie. Before that, I received my B.Sc. in Computer Science (ELITE Stream) from The Chinese University of Hong Kong, and my final year project was advised by Eric Lo. I also spent a semester as an exchange student at ETH Zürich during my undergraduate, where I explored advanced topics in Machine Learning and Software Engineering.
My research is supported by grants from Prof. Nie’s research group, funded by the Natural Sciences and Engineering Research Council of Canada (NSERC) and the University of Waterloo.
💻 Research Interests
I study how language models and agents reason, learn from experience, and use computation efficiently. My interests span four connected directions:
Efficient Inference
Making LLMs and agents more efficient by rethinking how they use computation, memory, and context, with a focus on scalable inference and serving for long-running tasks.
- KV-cache compression & eviction
- Agent-aware serving (vLLM / SGLang)
- Context & KV memory to weights
- Agent Efficiency
Latent, Looped & Agentic Reasoning
Understanding how LLMs reason beyond explicit CoT, and how learning and test-time computation can make reasoning more effective and reliable.
- Latent & looped reasoning
- Coding agents
- Agentic deep research
- Test-time scaling
LLM Architecture & Mechanisms
Investigating how model architectures and internal representations shape learning and reasoning, and exploring designs that balance expressive power, computational efficiency, and scalability.
- Tokenization
- Hidden representations
- Looped LLMs
- Hybrid attention
- Quantized latent states
LLM Agents & Test-Time Learning
Building agents that learn from experience, retain useful knowledge, and adapt their behavior over time, with a focus on sustained interaction with software and evolving environments.
- Self-evolving agents & harnesses
- Agent memory
- Computer-use agents
- Agents for ML engineering
Selected work
Architecture TokDrift
Agents OSWorld-Science · Learning from Use (survey)
I am open to academic and research collaborations and welcome any questions or discussions.
🎉 News
| Sep 29, 2026 | 📌 New preprint: OSWorld-Science: A Benchmark of Computer Use Agents for Learning and Using Scientific Software is now available on arXiv! Please check out the project page for more details! |
|---|---|
| Aug 20, 2026 | 📌 New survey: Learning from Use: Test-Time Learning in Large Language Models and Agents is now available! |
| Apr 30, 2026 | 🎊 One paper accepted to ICML 2026! |
| Apr 06, 2026 | 🎊 One paper accepted to ACL 2026! |
| Mar 02, 2026 | 🎊 One paper accepted to ICLR 2026 Workshop! |