Reproduce “A Survey on Knowledge Distillation of Vision-Language Models” 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
Hoang Tran Vuong, Dat Dao, Pham Khanh Chi, Minh-Phuc Truong, Quyen Tran, Linh Ngo Van, Thien Huu Nguyen, Thanh-Toan Do and 1 more
The authors survey knowledge distillation for vision-language models, organizing methods by transferred knowledge, teacher access, and structural compatibility. They explain why matching outputs may not align representations, and find that methods combining response and feature supervision score better in comparisons. On GQA, Switch-KD scores 63.9 versus 58.7 for LLaVA-MoD; confounds limit conclusions, and grounding-fidelity evaluation is lacking.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.