检索重排实测报告:候选过多时部分 reranker 质量不升反降
CShorten30 · x · 2026-11-18
CShorten30's team released the "Scaling Reranked Retrieval" research report; amplifier mrdrozdov highlighted two key findings:
- The "drowning in documents" problem persists: even recent rerankers sometimes drop in quality when reranking too many documents
- Listwise as a third retrieval stage looks compelling
Chart data: on five BRIGHT subsets, hybrid retrieval top-20 without reranking yields Recall@20 of just 0.282; ZeroEntropy zerank-2 with a k=2,000 pool raises it to 0.560; adding a Listwise GPT-5.6 Luna stage reorders the same 20 docs (Recall@20 pinned) and pushes Success@1 to 0.478.
所属事件:Weaviate 发布《Scaling Reranked Retrieval》研究报告并开源代码(10 条相关)→
「研究」频道最新
- 生成式 AI 进心脏病学:Dagstuhl 跨学科研讨会落幕 — StephanMandt · 2026-11-22
- SubRoformer 用 DSP 子带滤波器刷新音乐源分离 SOTA — affige_yang · 2026-11-22
- Emily Riehl 推出 Lean 证明入门游戏,用交互方式教依赖类型论 — abhi9u · 2026-11-22
- Android 移动 Agent 评测论文清单:从 AppAgent 到 AndroidReality — East-Muffin-6472 · 2026-11-22
- 研究者对比机制可解释性与神经科学:人类大脑如何做因果验证? — burny_tech · 2026-11-22
- 研究者提出“自然语言流形”概念,与物理时空相关词尤明显 — yasamanbb · 2026-11-22