Reranking more docs backfires: Recall@10 peaks near 100, then declines, study finds

CShorten30 · x · 2026-09-15

Weaviate Podcast #141 features Mathew Jacob's paper "Drowning in Documents": as a Databricks intern he found that scaling cross-encoder reranking depth lifts Recall@10 only up to 100 documents, after which it declines — a real, counterintuitive curve he initially mistook for a bug. The work proposes sliding-window listwise reranking (robust at 1,000 docs) and ranking cascades, and the episode also covers TraceLab, a dataset of 40,000 real coding-agent traces.

Original post →

More from coding & agent

coding & agent channel →