I'm Junjie 俊杰.
CS + Maths @ UC Berkeley
My research centers on reinforcement learning, training language model agents, continual learning, and LLM evaluation.
More broadly, I am interested in language models and machine learning systems, and I am currently deepening my understanding of performance engineering and efficient machine learning.
Outside of research, I am drawn to investing and the venture capital world, and I am an avid reader. I also enjoy staying active through swimming, long-distance running, and racket sports like tennis, squash, and pickleball.
Currently
I am looking for roles in post-training research, machine learning engineering (MLE), and software engineering (SWE).
Research
RL for language models and agents; advised by Raluca Ada Popa & Ion Stoica.
I am forever thankful for the mentoring from Sijun Tan, Tianhao Wu, and Sida Li.
Agents’ Last Exam (ALE). Building infrastructure for end-to-end evaluation of computer-use agents, with Dawn Song’s group and the SGLang team.
Efficient evaluation. Exploring early stopping, representative task sampling, and paired comparisons to evaluate language models and agents more efficiently while maintaining reliable performance estimates.
Neural data. Exploring pretraining and adaptation of language and diffusion models for neuroimaging and neural activity data.
NLP Group; advised by Dan Jurafsky and mentored by Chen Shani, exploring whether post-training can steer language models toward concept-preserving reasoning.
Contributed to the Vulci project with Maurizio Forte, doing geophysical surveys and computer graphics; mentored by Nevio Danelon and Antonio Lopiano.
Internships
Developed Go backend services and infrastructure for an internal ML platform supporting model training and LLM evaluation for Tencent’s Hy3 and Hy4 preview models. Contributed from day one to asynchronous workflow orchestration, automated failure recovery, and model proxy monitoring.
Explored methods for evaluating mathematical reasoning in LLMs, including auto-formalization of SPDE questions in Lean.
Collaborated with Xiaomi’s audio team on MiDasheng model development and product integration.
Supported deal sourcing and due diligence for consumer-sector investments, preparing research reports and tracking investment memos. Contributed to early-round investment work, primarily involving Chi Forest and Qiandama. Cowin was an early investor in Pop Mart, Chi Forest, Tea’stone, and Qiandama. Cowin is a private equity and venture capital firm founded in 2000, with approximately US$5.2 billion in AUM.
Honors
Dean’s List. In-class rankings: 1st in CS 168, 2nd in EECS 126, and 4th in EECS 127.