Jiachen Lian (连甲琛)

I am a PhD candidate in Berkeley EECS, affiliated with Berkeley Artificial Intelligence Research (BAIR), where I am advised by Prof. Gopala Krishna Anumanchipalli. I also collaborate closely with Prof. Maria Luisa Gorno-Tempini on precision healthcare for speech and language, and with Prof. Shankar Sastry on control theory, dynamics, and internal-model representations in AI foundation models. My work is supported by the Meta AI Mentorship (AIM) Program over two years, through which I have worked with Wei-Ning Hsu and the LLaMA 3 and LLaMA 4 teams, as well as Abdelrahman Mohamed.

Email  |  Github  |  LinkedIn  |  Twitter  |  Google Scholar  |  CV

I'm interested in research questions that will still matter five years from now.

I build interpretable, human-centered foundation models — and the systems and robots that bring them into the real world.

(1) Interpretable-by-design, mechanistically grounded foundation models
I develop foundation models whose internal representations are structured and inspectable, and whose reasoning, planning, and dynamics can be examined and controlled.

(2) Human-AI collaboration and deployment in high-stakes healthcare settings
I build collaborative, deployed systems for precision and public health — speech and language disorders such as dyslexia and primary progressive aphasia, and population-level language screening.

(3) Human-centered speech AI systems
I build speech AI that is grounded in human cognition and physiology (perception, production, physics) and that keeps people in the loop through interactive labeling, feedback, and supervision.

(4) Controllable robots and inverse problems
I work toward robots whose actions we can steer, predict, and understand.

Why do so few scientists make significant contributions and so many are forgotten in the long run [1]?
[1] https://www.cs.virginia.edu/~robins/YouAndYourResearch.pdf

• [Oct 2024] As approved by the California State Government, starting from 2025 public schools will adopt our language screener, where I developed the first and state-of-the-art speech dysfluency transcriber [UDM][SSDM], serving 1 million kids! See reports.

• [Oct 2025] Twice a year, we host the Bay Area Speech AI for Health, Education, and HCI Workshop, convening leaders from academia, industry, venture capital, healthcare, and education. Our goal is to shape the future of human-centered speech intelligence — connecting fundamental research with real-world systems and societal impact. We welcome researchers, practitioners, founders, and investors to join, present, or collaborate. Please feel free to contact me. https://sites.google.com/berkeley.edu/bair-speech-ai-workshop

• Full-Duplex-Bench: the default and most comprehensive benchmark for full-duplex conversational AI, adopted by Thinking Machines and Kyutai (Moshi).
   [GitHub]   arXiv: [v1] [v1.5] [v2] [v3]

• Core contributor to Meta's open-weight LLaMA models — total Hugging Face downloads:
   🤗 LLaMA 3.2 216,187,015  ·  🤗 LLaMA 3.3 12,902,451  ·  🤗 LLaMA 4 10,386,581