Welcome
Hello and welcome to my personal homepage! My name is Zichuan Fu, and I am thrilled to share my educational background and aspirations with you.
I completed my undergraduate studies at Xi’an Jiaotong University’s School of Software Engineering, where I earned my Bachelor of Engineering degree. I was fortunate to be mentored by Professor Jiakun Zhao during my time there. Following my undergraduate education, I pursued a Master of Science degree in Advanced Computer Science at the University of Sheffield. There, I had the pleasure of working under the guidance of Dr. Anton Ragni.
I am now excited to embark on my journey as a Ph.D. candidate at City University of Hong Kong, under the supervision of Professor Xiangyu Zhao.
As a passionate advocate for artificial general intelligence (AGI), I am eager to contribute my efforts to help bring this revolutionary technology to fruition. I firmly believe that AGI has the potential to reshape our world for the better, and I am committed to playing an active role in its development. Thank you for visiting my page and learning more about my background and aspirations.
Research
My work centres on making LLM agents reliable enough to trust with real tasks. Current agent frameworks hand the model the role of orchestrator, and then treat the resulting token explosion, control-flow hallucination, and unreliable stopping as bugs to be prompt-engineered away. I argue these are architectural consequences of asking a probabilistic system to do deterministic work, and that the fix is structural.
That line of work runs through three connected threads:
- Agentic Programming — putting control flow back in the program and invoking the model only where a task calls for judgement (LLM-as-Code, KDD’26 AgenticSE Workshop; implemented in OpenProgram).
- Grounded interaction — letting agents act in real interfaces through active visual observation rather than one-shot prediction (GUI-Lens; GUI-Agent-Harness, 79.8% on OSWorld Multi-Apps).
- Efficient and well-behaved attention — what models spend their context on, and why they misallocate it (Lazy Attention, SWAT).
Alongside these I work on model merging and knowledge editing, temporal and multi-hop reasoning, and LLMs for recommendation. Code for most projects is on GitHub.