Collaborations

TL;DR

If you think I can be helpful for your project or you are looking for a project. Feel free to shoot me an email.

For undergraduate students

I have some ideas on projects in AI and Systems. They are in different directions and have different levels of maturity. Previous collaborations with more junior researchers have resulted in paper like HypoEval. If you are interested, feel free to shoot me an email with your background and your interest.

Direction that I am currently working on:

LLM inference. If you want to help out. Also shoot me an email with your background.

Some other questions that I think about:

• Non-volatile memory devices are getting very fast (>10GB/s). Can we improve AI/multimdeia applications utilizing these hardware/system innovations? I think this is really interesting but I do not have a full project in mind. Here is a list of some related works on this link.

• Post-training has demonstrated significant gains in LLM performance. How can we:

  1. collect preference data during serving?
  2. design new metrics outside preference to optimize other things like retention rate?

  The first question is due to cost of collecting annotation. It is ideal if we can make self-improving LLMs. I have an initial design about such a system.

  The second question is more about designing metrics that application builders eventually care about. For example, if the only profit source is from ads, there will be a need to optimize ads click rate and preference at the same time.

• In reasoning models, models will need to iterate through many options before getting the right onw. Can we compress the reasoning chain after thinking finished? Note that this is different from the "overthinking" since LLMs will still need to think about many things before realizing it is the wrong direction.

• When we chat with human beings, we adjust our attitude and speaking style based on the other party's feedback. Can (or should) we design a system that tells if the user is impatient and adjust model behavior? (ex. through adding some hint prompts like "user is getting impatient, I need to be more concies." )