Hi! I’m Juncheng Liao (廖俊程), a master’s student in Software Engineering at Zhejiang University. I previously studied Intelligent Vehicle Engineering at Harbin Institute of Technology.

My interests sit at the intersection of multimodal large language models, efficient model training and inference, and continual learning. I enjoy turning research ideas into reliable systems and contributing the resulting improvements back to open-source projects.

Education

Zhejiang University · Master’s in Software Engineering
School of Software Technology · Sep 2025 – Jun 2028 (expected)

Harbin Institute of Technology · B.Eng. in Intelligent Vehicle Engineering
School of Automotive Engineering · Sep 2021 – Jun 2025

Research Interests

• Multimodal large language models
• Continual learning and catastrophic-forgetting mitigation
• Efficient training, inference, and low-precision kernels

Selected Work

• AWARe — explored activation-weighted parameter importance to mitigate catastrophic forgetting in multimodal large language models.
• Open-source systems — contributed FP8 GEMM support for NVIDIA Blackwell GPUs to vLLM and SGLang, video rollout support to VeRL, and GLM-4.1V training support to Transformers.