Guang Liang (梁广)

Guang Liang (梁广)

PhD Student at Nanjing University (LAMDA Lab) & Beijing Zhongguancun Academy. Research: Efficient LLM Training/Inference, Model Quantization.

Hi, I am Guang Liang (梁广)

I am a PhD student at Nanjing University (LAMDA Lab, advised by Prof. Jianxin Wu) and Beijing Zhongguancun Academy (advised by Prof. Tie-Yan Liu).

Previously, I received my B.Eng. from Xi’an Jiaotong University (AI Honors Class, Qian Xuesen College) in 2024.

🔬 Research Interests

My research focuses on efficient and generalizable AI systems, particularly:

  • High-Efficiency Model Design: FP8/FP4 training, quantization-aware training (QAT), and optimization algorithms for better data efficiency. Goal: achieving higher intelligence with the same training compute.

  • Efficient Training & Inference for LLMs: Pretraining, mid-training, SFT, and RL for LLMs and VLMs. Focus on faster training (FP8, 1000+ GPU clusters), better generalization, and lower deployment costs (quantization & distillation).

💼 Experience

  • Shanghai AI Laboratory - OpenDATALab Research Intern (Supervisor: Dr. Conghui He, Mentor: Dr. Bin Wang)
    • Led architecture innovations for DocLayout-YOLO (2k+ ⭐)
    • Core contributor to MinerU (55k+ ⭐): formula recognition & efficient VLM training
    • Co-first author of UniMERNet (450+ ⭐): proposed R-S attention mechanism

🏆 Selected Honors

  • Champion (Best Solution Award, ¥200,000) - MVA 2025 Challenge: SMOT4SB (1st/78 teams, 308 submissions defeated)
  • Champion (1st Prize, $1,000) - TRAUMATHOMPSON@MICCAI 2023 (Ranked 1st in both Action Recognition & Action Anticipation tracks)
  • Reviewer: CVPR 2026, ECCV 2026

📫 Contact

📄 CV

Download my full CV (PDF)