Reproduce “Calibrating Document-Local Boundaries for Contextual Chunk Retrieval” 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
Yuanxiang Wang, Hui Jiang, Da Zhu, Zhihua Chang
DcLBC addresses same-document chunk confusion in contextual retrieval. It combines Adaptive Margin Scaling, which focuses training on each query’s hardest sibling, with Chunk-Aware Key Routing, which lets chunk boundaries influence attention during encoding. On ModernBERT-large, it raises ConTEB’s eight-task macro NDCG@10 from 76.70 to 78.61, and NanoBEIR from 59.75 to 60.56; improvements are not uniform across tasks or backbones.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.