Hi! I'm Ziche Liu, a first-year Ph.D. student in Computer Science and Engineering at UC San Diego, where I am fortunate to be advised by Prof. Jingbo Shang. I received my B.S. from the Chinese University of Hong Kong, Shenzhen, and spent a year at UC Berkeley as a visiting student during my undergraduate studies.
Previously, I worked closely with Dr. Feng Jiang under the supervision of Prof. Haizhou Li on data selection for LLM fine-tuning, and with Prof. Benyou Wang on model localization and bias analysis in LLM-as-a-judge. During my time at UC Berkeley, I also worked on world models at Berkeley AI Research (BAIR).
My research interests broadly lie in LLM, agents, and embodied world models, particularly in understanding, evaluating, and improving their ability to learn, reason, and interact with complex environments. I'm always happy to connect and chat about research so please feel free to reach out!
I work at the intersection of NLP/LLMs and embodied world models, asking how data, architectures, and evaluation shape what models really learn.
Language & LLMs. My NLP work has studied which data actually helps large models, how human and LLM judges can be biased, and how to adapt systems across cultures. I think of language as a compact interface on top of richer perception and interaction, not a full replacement for them.
Embodied world models. At BAIR, I build action-conditioned video world models that let agents imagine futures and respond to control signals over long horizons. More broadly, I'm interested in models that learn language, vision, and action together, instead of treating one as a small add-on to the others.
Evaluation & learning over time. I also work on how to evaluate these grounded agents in interactive, changing settings, where they need to handle uncertainty, ask good questions, and keep learning without forgetting.
Broader curiosities. I'm intrigued by human memory, consciousness, and intelligence, and how they might quietly inspire the next generation of grounded models.
TL;DR: We introduce RobotAgentBench, a benchmark spanning 50 household layouts that evaluates robot agents on task discovery, contextual reasoning, manipulation, failure recovery, and robustness, from individual tasks to open-ended scenarios.
TL;DR: We introduce SciLaws-Bench, 118 literature-grounded problems across six disciplines that test whether LLMs can discover scientific laws from real data and recover newly synthesized hidden laws in simulated parallel worlds.
TL;DR: We propose a three-stage scheme to standardize data selection methods and develop two metrics (efficiency and feasibility) to evaluate the effectiveness of a data selector.
When I'm not coding, you'll probably find me:
Yes, this is the place where people drop some fun projects: