Reranking isn't dead: cross-encoders saturate nDCG@10 but boost recall pools for LLM listwise rerankers
CShorten30 · x · 2026-10-08
CShorten30 and bowangbo debate whether reranking is dead.
- Cross-encoders collapse at @1 and saturate nDCG@10, but retain strong recall@20/50/100 pools that beat the first-stage retriever — handing those pools to an LLM listwise reranker performs much better than inheriting the retriever's top-100
- High-latency/higher-quality search still has use cases: agent trajectory robustness (Deep Research-style), search experiences, and search quality monitoring / training data curation
- bowangbo's evals: 30% of cases cross-encoders lift performance, 40% add nothing, 30% ruin the first-stage retriever; zerank-4b and Qwen-8b rerankers disappointed, with weak OOD confidence
- Side note: their reranking paper has been stuck in arXiv's review queue for over a month
Related event: Debate: Are Cross-Encoder Rerankers Dead?(4 posts)→
More from coding & agent
- OpenClaw slashes 1.3M flaky tests, doubling CI speed in big cleanup push — heyneighbor · 2026-10-09
- AI repos lost half their human code review since 2024, analysis of 2000+ projects finds — ronbodkin · 2026-10-09
- Thesis-defense audits proposed as gate for AI-generated code contributions — ronbodkin · 2026-10-09
- How do you decide what to optimize first in production AI agents? — Successful-Ask736 · 2026-10-09
- Matt Pocock's "Skills for Real Engineers" repo hits 280k GitHub stars — adnan_hashmi · 2026-10-09
- Graded thesis-defense reviews proposed as release gate for AI code — ronbodkin · 2026-10-09