A new arXiv paper asks which properties of hard-negative supervision can predict dense-retriever transfer before training. The authors compare lexical noncoverage, frozen-encoder similarity statistics, hardness proxies, and combinations of these signals.

The evaluation is carried out across BM25, a standard lexical retrieval baseline. The goal is to identify signals that track transfer without requiring expensive training runs.

Because the abstract is truncated, the article does not include results or conclusions. The study's contribution is in framing the question and proposing a comparison of pre-training signals.