Liang Chen

Liang Chen

Postdoc, EPFL

About me

I am a postdoctoral researcher at EPFL, working with Prof. Antoine Bosselut. I received my Ph.D. from The Chinese University of Hong Kong, advised by Prof. Kam-Fai Wong, and was a visiting researcher at LMU Munich with Prof. Hinrich Schütze. Earlier, I received my M.S. from Peking University and B.S. from Northwestern Polytechnical University.

My research goal is to make language models more capable without making them less reliable. In post-training the supervision is always imperfect, yet standard recipes are designed as if it were clean. I work on understanding how that imperfection surfaces in the trained model, and on building training methods that keep models capable and reliable in spite of it.

  • Learning from a sparse RL signal. In practice, RL for LLMs gives sparse rewards at the end of a long trajectory. Through denser credit assignment and wider exploration, we make it a stronger learning signal. [BRIDGE ICML’26, EEPO ICLR’26]
  • Making alignment survive what comes after it. LLM alignment is fragile: a range of downstream operations can undo it, from adversarial inputs to malicious finetuning. We build robust alignment methods that hold up under such perturbations. [PEARL ICLR’25, VAA ICML’25]
  • Trusting text the model produced. Post-training now runs largely on the model’s own output. We asked whether that output can be trusted — evaluating LLMs as knowledge generators, and making what they write attributable. [WatME ACL’24, CONNER EMNLP’23]

Feel free to reach out if you’d like to chat about research or explore potential opportunities.

News

Selected publications (Full list)

Talks

  • Learning Good LLMs from Imperfect Data
    Invited Talk, Microsoft Research Asia – November 2025
    Host: Dr. Jing Bai

  • Learning Good LLMs from Imperfect Data
    Invited Talk, EPFL – October 2025
    Host: Prof. Antoine Bosselut

  • Beyond Two-Stage Training: Cooperative SFT and RL for Improved LLM Reasoning
    PhD Seminar, LMU Munich – August 2025
    Host: Prof. Hinrich Schütze

  • Vulnerability-Aware Alignment: Protect Open-Source LLMs against Unsafe Fine-tuning
    AI Time, Online Live – June 2025

  • Towards Trustworthy LLMs: Improving Robustness via Post-Training Optimization
    PhD Seminar, LMU Munich – May 2025
    Host: Prof. Hinrich Schütze

Teaching

Teaching assistant at CUHK:

  • Operations Research II (SEEM3440)
  • Engineering Innovation and Entrepreneurship (SEEM3450)
  • Information Technology Management (SEEM5730)

Internships

  • Microsoft Research, Systems Research Group
  • Tencent AI Lab, Machine Learning Center

Community service

  • Area Chair for ACL ARR.

  • Program Committee for AAAI 2027.

  • Reviewer for ICML, ICLR, NeurIPS, TMLR, AISTATS, ACL, EMNLP, and NAACL.

Miscellaneous

Outside of research, I enjoy walking in parks, as well as swimming, hiking, and table tennis.

During my undergrad, I was the runner-up in the Freshmen Cup table tennis singles match and won the team championship three times.

Live long enough to live forever