Reproduce “RandSlot: Learning Compact Visual Document Representations with Random Soft Tokens” 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
Dewen Guo, Shi Yu, Lingxiao Zhang, Yang Zhang, Tao XU, Dan Wang
RandSlot addresses compact visual-document retrieval by appending independently resampled, normalized random positions during training and using their contextualized encoder outputs as readout vectors. A shared causal encoder produces four vectors for each query and page, scored with late-interaction MaxSim. On ViDoRe V3, Qwen2.5-VL-3B RandSlot averages 45.4 nDCG@5 versus CausalEmbed’s 42.6, but remains below full-token retrieval at 50.2.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.