Reproduce “MCD: Causal Distillation of Multimodal In-Context Learning in Large 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
Yanshu Li, Jiaqian Li, Canran Xiao, Xi Xiao, Tianyang Wang, Yongtai Liu
MCD distills how a vision-language teacher uses image-text demonstrations in few-shot question-answering prompts into a smaller student. It ranks tokens with teacher gradients and tests selected evidence by comparing outputs when evidence is retained or replaced. Across three model families and seven benchmarks, MCD improves scores by 7.23 points over original students and 4.68 over vanilla distillation in same-family settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.