Banghua Zhu
 |
Incoming Assistant Professor of Electrical and Computer Engineering at the University of Washington, |
About me
I'm an incoming assistant professor at UW ECE, with adjunct appointments in CSE. I received my PhD from the Department of EECS, UC Berkeley. I am very fortunate to be advised by Prof. Jiantao Jiao and Prof. Michael I. Jordan. I'm a recipient of the 2023 David J. Sakrison Memorial Prize from Berkeley EECS for truly outstanding PhD research.
I co-founded Nexusflow AI since 2023. We are dedicated to provide reliable LLM solutions for enterprise use-cases. Please feel free to contact me if you are interested in collaborations!
I'm recruiting PhD students in the 2024-2025 cycle! Please mention my name in the application of UW ECE / CSE if you'd like to work with me on LLM, game theory and / or reinforcement learning!
Research
I'm currently interested in the training, serving, evaluation, and applications of foundation models. I have also been working on statistics, information theory and machine learning, with applications on game theory, robust statistics, reinforcement learning and human-AI interactions.
Training Large Langugage Models
Checkout our open 7B model, Starling-7B, which ranks first in all existing Mistral-based 7B models according to human evaluation in Chatbot Arena! Starling-7B is trained with our open-source high-quality preference dataset, Nectar, using our new reward training and policy-finetuning algorithms.
We also released Athene-70B, currently ranking 11th among all models on Chatbot Arena. As a model fine-tuned from Llama-3-70B, the ranking increases from 22th to 11th on Arena, through our unique data curation and post-training pipeline.
We have open-sourced NexusRaven-V2-13B, a strong function-calling model that surpasses GPT-4 in complex function calling tasks, especially in terms of parallel and nested calls.
Evaluation
We have released Huggingface function calling leaderboard, which is used in Llama-3.1 technical report for evaluating function calling capabilities.
Chatbot Arena is one of the most reliable platforms for evaluating model with human preferences.
Arena-Hard-Auto is an automatic benchmark creation pipeline that uses LLM-as-a-judge to quickly evaluate the model performance.
Main research Interests:
Foundation models, or specially large language models (LLMs).
Finetuning and Evaluation of LLMs: In my research on reinforcement learning with human feedback (RLHF), I identify the fundamental limit and develop near-optimal algorithms with improved sample complexity for reward training [ZJJ23]. I also propose alternative of Proximal Policy Optimization (PPO) for policy optimization that is more stable and sample-efficient [ZSFDZJJ23]. These advancements enable more efficient and reliable model fine-tuning.
Inference in LLMs: I analyze and propose near-optimal algorithms for caching and model multiplexing for serving large models, significantly enhancing the efficiency of inference in LLMs [ZSZBJJ23].
Human-AI interactions.
ML w/ Helpful Human: I focus on designing advanced human-in-the-loop AI systems that learn from and collaborate with humans. Using a theoretical framework based on Stackelberg game theory, my research explores how AI benefits from interactions with human experts, contributing to a deeper understanding of these domains [ZZJJ23].
ML w/ Strategic Human: My research explores the interaction between machine learning systems and self-interested, strategic humans—a crucial topic in economics. By modeling and analyzing online learning in contract theory and the creator economy, I provide near-optimal regret bounds for both problems, addressing the longstanding challenge of sample complexity in online contract design [ZBYWJJ23, ZKJJ23].
ML w Malicious Human: I explore techniques to enhance the resilience of AI models against malicious attacks. I extend the theory in high-dimensional robust statistics [ZJS22], and propose efficient algorithms for outlier detection, robust mean estimation, robust covariance estimation, robust linear regression [ZJS21], and Byzantine-robust distributed learning distributed systems [ZPWWJSJ23].
Besides the topics above, I am also actively working on the following research areas:
Bandit and Reinforcement Learning: I study online learning and offline learning, off-policy evaluation and inverse RL [RZMJR21, MZJW22];
Information-theoretic Lower Bounds: I investigate achieving fundamental limits in noisy searching, sorting, and computing tasks using information-theoretic tools. [WGZW22, ZWGJW23];
Semi-Supervised and Unsupervised Learning: I design doubly-robust estimators that outperform traditional self-training pipelines in computer vision and autonomous driving [ZDJWZJJ23]. Furthermore, I conduct theoretical analyses of Generative Adversarial Networks (GANs), providing insights for practical implementations [ZJT19].