Reproduce “SmallReason-ColBERT: An Ultra-Small Late-Interaction Retriever for Reasoning Intensive 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
Abdelrahman Abdallah, Mohammed Ali, Adam Jatowt
SmallReason-ColBERT adapts a 32M-parameter ColBERT retriever for reasoning-intensive queries. It retrains a compact model on reasoning-aligned data and adds a 129-parameter query-token weighting head, trained on a frozen base. The head weights token matches using distinct training and evaluation scores. It reaches mean nDCG@10 of 21.41 across BRIGHT’s 12 splits, exceeding evaluated compact baselines but retaining late-interaction costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.