Querit-Reranker Paper Pushes BEIR nDCG@10 from 54.11 to 59.28
Jayyun98_dev · x · 2026-10-09
Responding to "R.I.P Reranker" claims, Jayyun98dev cites the arXiv paper Querit-Reranker:
- A data-centric, label-free adaptation pipeline for compact multilingual rerankers: synthetic-query mining with teacher scores as continuous soft labels, plus spherical-linear-interpolation checkpoint merging into a single deployable model with no runtime ensembling.
- Two instances: Querit-Reranker-A0.4B (MoE backbone, 0.4B activated params) and Querit-Reranker-4B (initialized from Qwen3-Embedding-4B).
- With Qwen3-Embedding-0.6B as first-stage retriever, A0.4B lifts average nDCG@10 from 54.11 to 59.28 on BEIR and 59.87 to 67.70 on MIRACL.
The author argues cross-encoders still deliver real gains on the same retriever's top-100 and questions whether alternatives match that quality at lower end-to-end cost.