About

I am a 2nd-year PhD student at Stanford NLP advised by Diyi Yang. I had the pleasure to work with Monica S. Lam and Michael Bernstein during the rotation program. Previously, I was an undergraduate student at Yuanpei College in Peking University where I got into ML and NLP research by working with Bing Liu. In summer 2022 , I had a research internship in UCLA hosted by Nanyun Peng. Before that, I have worked as a research intern in Microsoft Research Asia (blog spotlight in Chinese) and an engineering intern in Tensorflow Lite team at Google, Beijing.

My research interests lie in ML and NLP. Nowadays, I’m interested in positioning NLP models (e.g., LLM) into larger systems. Here are some core problems I’m thinking about:

How can AI models bridge human and systems or systems and systems?
How can AI-empowered systems collaborate with users effectively?
How to continually improve these systems through the interaction with human and external systems?

Many kind people helped me a lot in my journey. If you want to talk more about research or seek advice that I might be able to provide, feel free to book a chat here.

News

(Jan, 2025) Collaborative Gym, a framework for enabling and evaluating human-agent collaboration, is out on arXiv. Check out our preprint and Twitter thread! Working on making collaborative agents accessible and codebase release 💪, stay tuned!
(Jan, 2025) Will give a talk on LM Agent Privacy Risk (paper) at JP Morgan AI Research seminar.
(Sep, 2024) PrivacyLens is accepted to NeurIPS 2024 Datasets and Benchmarks Track. Super excited to attend NeurIPS for the first time!
(Sep, 2024) Gave a guest lecture on "Knowledge Curation" at CS 224V with Yucheng and Prof. Lam. You can find the slides here.
(Sep, 2024) Collaborative STORM, a major update to the STORM project that brings human in to the loop, is now accepted to EMNLP 2024. Our knowledge-storm package is now v1.0.0! [release note]

Selected Projects

LM-Empowered System for Knowledge Curation

We study the development of knowledge agent for writing long, organized, and well-grounded articles, and how humans can collaborate with knowledge agents.

Highlights:

Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models, In NAACL 2024 (📰 Wikimedia Coverage)
We introduce STORM, a system that generates full-length, Wikipedia-style articles by conducting perspective-driven literature research and organizing information into structured outlines.
Into the Unknown Unknowns: Engaged Human Learning through Participation in Language Model Agent Conversations, In EMNLP 2024
We introduce Collaborative-STORM, a human-in-the-loop multi-agent system designed to uncover an individual’s “unknown unknowns” within their topics of interest by scaffolding human participation in multi-agent conversations
storm

Continual Learning in NLP

We study (1) continual pre-training/post-training of language models (LMs) and (2) enabling LMs to continually learn new tasks after deployment.

Highlights:

Continual Pre-training of Language Models, In ICLR 2023
We propose a post-training algorithm with adaptive soft-masking mechanism that selectively updates LM parameters based on the post-training corpus to minimize catastrophic forgetting and enhance knowledge transfer.
Class-Incremental Learning based on Label Generation, In ACL 2023
We investigate continual learning with classification objective and generation objective by examining representation collapse in pretrained models throughout the learning process.
ContinualLM