About Me

I am a Master’s student in Artificial Intelligence at Huazhong University of Science and Technology (HUST). My research focuses on trustworthy machine learning.

Research Interests

• Out-of-Distribution (OOD) Detection
• Test-Time Adaptation
• Trustworthy Machine Learning
• Deep Learning

Education

• 2025.09 – Present, M.S. in Artificial Intelligence, Huazhong University of Science and Technology
• 2021.09 – 2025.06, B.S. in Artificial Intelligence, Huazhong University of Science and Technology

Future Work

Thanks to the digital infrastructure built over the past 40 years, coding has enabled LLMs to take the first step toward self-evolution, simply because it comes with compilers and unit tests. The model writes code, and the compiler and test cases tell it right or wrong immediately. That forms a perfect “generate–verify–revise” loop, which is a crucial kick-off stage toward AGI – it’s the first time we’ve seen just how massive the potential of LLMs really is.

Looking ahead, no matter whether the underlying architecture is a variant of Transformer, State Space Models (SSM), or some brand-new paradigm we haven’t named yet; no matter whether alignment uses RLHF, DPO, or fancier stuff like Constitutional AI – this current training paradigm, driven by rigid rules and heavy compute, will keep pushing the reasoning foundation of models for a long time to come, making them approach human-level performance in deterministic domains like math, programming, and logical deduction.

This rule-based self-evolution hasn’t hit its ceiling yet, but we all know that when a single metric becomes the only quantifiable target, it stops being a good measure. Today’s systems can answer questions perfectly, yet they never show the curiosity of a baby toward the unknown – they’re always just waiting to predict the next token, never feeling the boundary of their own predictions. Multi-agent collaboration and competition do reveal some interactions at the boundaries of capabilities, but for now, such designs still rely on preset frameworks.

That’s why my focus has shifted to the safety boundary and trustworthiness of AI systems. Its value is not just about preventing harmful outputs; it’s about laying the research groundwork for a broader sense of self-evolution. Trustworthiness research shouldn’t just teach a model what it’s not allowed to say – it should give it a sense of its own limits. A model that can sense its own uncertainty can actively fill in the gaps, and then continuously recalibrate itself, rather than passively waiting for data from humans.