Yuheng Tu

I am a fourth-year undergrad from School of Information Science and Engineering, Southeast University. I was a visiting student at UC Berkeley EECS during 24 Spring.

I had the privilege of working with Prof. Bo Li and Prof. Sanmi Koyejo during 24 Summer. Before this, I was fortunate to be advised by Prof. Qiao Wang at Southeast University. Additionally, I had the opportunity to collaborate with Prof. Yinfei Xu and Prof. Jianan Zhang on various research projects.

I am a CS self-learning advocate and my study notes are presented in CS-self-learning.

Email  /  CV  /  Statement  /  Scholar  /  Github

Research

My experience spans various areas within Machine Learning, with a recent focus on AI Evaluation and Trustworthy AI. I am broadly interested in drawing insights from established theories to advance AI development, while exploring the new life of traditional methods in the context of contemporary AI.

AI Evaluation with statistical model and question generator.
Evaluate 184 LLMs across 25 datasets reliably and efficiently with Item Response Theory (IRT). Introduce amortized calibration to reduce the cost with minimal sacrifice of accuracy. Fine-Tune Llama-3-8B to generate questions conditioned on item parameters.

Safety benchmark aligning with regulations and policies.
Generate 5,694 detailed and diverse instruction prompts across 314 risk categories and 3 language styles. Evaluate 22 leading LLMs with GPT-4o as a judge and category-specific system prompts.

Gradient compression problem in Federated Learning.
Develop the NQFL algorithm which normalizes the gradients and quantizes them with the Lloyd-Max quantizer. Implement NQFL along with 3 comparative algorithms: QSGD, AdaQuantFL, and SLMQ in FedML framework.

Model structure problem in Supervised Regression.
Develop the neuro-TF model (combines neural networks and pole-residue-based transfer function) which provides accurate and fast prediction of the EM behavior of a metasurface and thus greatly accelerate the design process.