DaRL Group / Hua Wei

In a recent DaRL Lab group meeting, I presented a technical breakdown of the underlying architecture of Claude Code. As a terminal-based AI agent tool, Claude Code addresses engineering challenges regarding real-time interaction, extensibility, and state management within large context windows.

AI is not replacing teachers. In many ways, it is allowing teachers to return to the true essence of the profession: understanding students, supporting them, helping them build confidence, and guiding them as they find their direction.

AI in My Classroom

I started integrating AI more actively into my teaching this academic year. During class, I ask students to use the NotebookLM and Gemini agents I configured for the course. These agents allow students to ask questions about…

How to port n-step returns, TD(λ), uncertainty, safety, and friends from deep RL into RLHF/RLAIF and tool-using LLM agents — with equations.
 

This note collects well-known deep RL techniques (pre-LLM era) and adapts them to modern LLM training/inference. The patterns below are not the only way to do it — treat them as practical starting points you can refine for your stack.

Why bring deep RL ideas into LLMs?

A generation is a trajectory:…

When I was preparing my CAREER proposal, especially the first time, I often felt completely overwhelmed. Every day was a tug-of-war between “I cannot look at this anymore” and “Wait, maybe I can still improve this part…” But even then, I was never quite sure where to focus, or what the reviewers were really looking for.

Fortunately, along the way, I came across many helpful resources – traces left behind by those who had been through it. These really…

This post is immigrated from the original page.

[paper] [code] [video]

As one of the most popular technologies in the field of machine learning in recent years, reinforcement learning (RL) has made significant progress in areas such as games, robotic control, and large model training. It feels like RL is back!!!

However, when we talk about its practical applications in the physical world, there are still many challenges to overcome—beyond the classic issues like learning efficiency and generalization, real-world deployment requires solving even more…

Starting as a researcher in spatio-temporal data mining, I still vividly remember the excitement of my first paper being accepted by CIKM 2016 on predicting passenger demand for car-hailing apps. In the last 7-8 years, the field of traffic prediction has gained substantial attention in prominent conferences like KDD, AAAI, and IJCAI. However, as time goes by, a recurring theme in this topic has been the concern…

https://darl-lab.gitbook.io/handbook/

Or Chat with AI Hua Bot here (ASU Login Required).