Siqi Li

Hi! I’m Siqi 👋

I work on vision-language-action (VLA) models for robotics, with a focus on dynamic manipulation—tasks where precise timing, velocity, and physical interaction are essential to success.

My research asks a simple but practical question:
how do we make VLA models work well in dynamic tasks — moving objects, changing scenes, unexpected interruptions — the kind of conditions real applications actually demand? I aim to extend their capabilities beyond controlled benchmarks to meet the demands of real-world applications.

I combine theoretical analysis with hands-on systems engineering, translating research ideas into working robotic systems and evaluating them in simulation and on real hardware. My approach pairs practical solutions with principled analysis to understand why they work, where they fail, and how to improve them.

I’m a PhD student in EECS at UC Irvine, advised by
Prof. Yasser Shoukry in the
Resilient Cyber-Physical Systems Lab.
Previously, I was a visiting researcher at Caltech, working with
Prof. John Doyle on language-to-robot control.

LLM Personalization

Alongside robotics, I work on making large language models fit each individual user — and keep fitting them as the user and the model interact more and more over time.

The vision is simple: every user keeps a local, private preference memory on their own device — a batch of their past choices, feedback, and preference statements that stays on their end rather than in the model. At inference time the LLM retrieves from that memory, RAG-style, to understand who it is talking to and generate preference-aligned responses, without retraining and without shipping personal data to a server.

Along the way I study questions like:

• which signals in a user’s history are actually stable and informative (e.g., personality traits), and which are noise
• how to retrieve the right memories for a given query rather than simply the most similar ones
• how to evaluate whether a personalized response is truly aligned with the user, not just plausible

The goal is an LLM that becomes more yours the longer you use it — while the data that makes it yours never leaves your hands.

A Bit More About Me 🌍

I work fluently in three languages (Mandarin, English, and German), both conversationally and academically, and I’ve lived and studied across Asia, Europe, and the US.

I enjoy travel, food, and building things that work — which is probably why I’m drawn to problems where theory meets reality.

I believe the most interesting AI research is:

• rigorous but not fragile
• principled but hands-on
• serious, yet a little fun

If that resonates, feel free to reach out — I’m always happy to chat.