Computer Science undergraduate at the University of Wisconsin–Madison.
Hi, I’m Moyang, also known as Steven. I am interested in how we train, adapt, and control language models, and how to make their computation more efficient.
My research interests include representation editing, activation steering, parameter-efficient adaptation, and efficient inference. I am also interested in LLM post-training and multimodal foundation models.
In my current independent research, I study the relationship between model interventions and cached computation. I build experiments around LoRA and KV caches to distinguish changes in model parameters from effects carried by prior computation.
I enjoy working from the underlying math to a concrete implementation, then testing where an idea works and where it breaks. Outside research, I practice calisthenics and gymnastics rings.
Feel free to get in touch to discuss research or potential collaborations.
When an intervention changes a model, what changes in its parameters, activations, and cached history? I build controlled experiments around LoRA, activation steering, and KV caches to separate these effects.
My current work investigates when cached computation can be reused or corrected under model adaptation, and how to evaluate the quality–compute trade-off.
Research overview and experimental lessonsAt Beijing University of Technology, I contributed to a research proposal on building a course-specific knowledge graph. The design combines a domain ontology, educational-text processing, entity recognition, and graph storage.
The documented proposal specifies BiLSTM-CRF and Neo4j; its literature review includes BERT-based entity prediction.
Outside computer science, I enjoy calisthenics and gymnastics rings. I like the patient, incremental process of learning a difficult movement.
Away from the keyboard.