Reproduce “ProQuant: Progressive Quantization-aware Training for Edge MLLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yufei Xue, Yushi Huang, Jiawei Shao, Pingcheng Dong, Yonghao Tan, Shiyao Li, Kwang-Ting Cheng, Jun Zhang
ProQuant trains multimodal models for low-bit deployment using self-relabeling to match training answers to the target model and progressively quantizing vision-sensitive layers first. At W4A8, Qwen3-VL-2B-Instruct reaches 72.54% average accuracy across seven benchmarks, versus 65.94% for adapted LLM-QAT and 73.63% for BF16. At W4A4, it scores 59.83 versus 27.29, showing gains over the baseline but lower accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.