WorkRB exposes blind spots in skill benchmarks

This is a strong case for evaluating recommendation quality across the entire ranking window, not just whether known positives appear near the top.

• –Graded nDCG@100 sharply separates systems that find skills from systems that avoid irrelevant ones.
• –The challenge’s 154 valid submissions largely converged on multi-stage retrieval-and-reranking pipelines.
• –The winner scored 0.8458 versus 0.6292 for the strongest baseline, but used roughly 127 times more parameters.
• –Results suggest binary supervision can actively teach models to bury useful near-matches and tolerate lexical shortcuts.
• –WorkRB’s methodology could transfer to other recommendation tasks with sparse or incomplete annotations.