Most rerankers hurt retrieval: gains only show up at 4B scale, per new testing
bo_wangbo · x · 2026-10-17
Three empirical findings for retrieval pipelines: (1) bge-m3 is a strong embedding, but bge-reranker and many other rerankers mess up the first-stage retriever; (2) reranker gains only become visible when scaling the model to 4B; (3) the author reused pplx-decider — never trained on reranking supervision — with a prompt asking for 0-9 relevance ratings, evaluated on an in-house MTEB-style bench.
More from Research
- MajorTOM:Elliot Pretrain: 279K multi-sensor, multi-temporal tiles for Earth observation models — HildeKuehne · 2026-10-17
- Gen-Verse Releases DiscoveryRSI-27B for Recursive Scientific Hypothesis Development — jacek2023 · 2026-10-17
- AI agents hide mistakes in their own reports — GPT-5.5 omitted negatives in 198 of 200 abstracts — The Decoder · 2026-10-17
- Official vs open benchmarks disagree wildly — a call for an open calibration benchmark — lahwran_ · 2026-10-17
- Diffusion Explainer visualizes Stable Diffusion's noise-to-image process step by step — techNmak · 2026-10-17
- "Just add more RL envs": the one-line mantra of the RL scaling era — burny_tech · 2026-10-17