Yixiao HUANG
Berkeley, CA, United States
yixiaoh@berkeley.edu My first name can be pronounced as “e-/ɕjɑʊ/”:
I’m Yixiao HUANG (Chinese name: 黄一笑) and I’m a second-year PhD student at UC Berkeley EECS supervised by Prof. Somayeh Sojoudi. I also work closely with Prof. Song Mei. I’m interested in establishing the theoretical foundations for generative models to enhance their efficiency and reliability.
I received my Bachelor’s degree in Informaton Engineering in June, 2023 from Department of Electrical Engineering, City University of Hong Kong. I was fortunate to work with Prof. Rosa CHAN who patiently guided me to start my research adventure. I also had a great time at University of Michigan, working with Prof. Samet Oymak on the theoretical foundations of self-attention and hybrid models.
news
| Jun 12, 2026 | Introducing Agent’s Last Exam, a large-scale benchmark evaluating AI agents on long-horizon, economically valuable professional tasks. Check out what we find on the importance of the harness vs. models, and how we confirm the downgrade of Claude Fable 5. |
|---|---|
| Jun 12, 2026 | Two papers are accepted by ICML 2026: |
| Jun 12, 2026 | A new preprint on implicit Chain-of-Thought is out! |
selected publications
* indicates equal contribution- PreprintTransformers Provably Learn to Internalize Chain-of-ThoughtarXiv preprint arXiv:2605.28600, 2026
- ICML 2026Breaking the Reversal Curse in Autoregressive Language Models via Identity BridgeInternational Conference on Machine Learning (ICML), Spotlight, 2026
- NeurIPS 2025Generalization or Hallucination? Understanding Out-of-Context Reasoning in TransformersAdvances in Neural Information Processing Systems (NeurIPS), 2025
- NeurIPS D&BOVERT: A Benchmark for Over-Refusal Evaluation on Text-to-Image ModelsAdvances in Neural Information Processing Systems Datasets and Benchmarks Track (NeurIPS D&B), 2025