Yuanbao-Reranker：文档接地合成与偏好蒸馏的文本重排序
原标题：Yuanbao-Reranker: Document-Grounded Synthesis and Preference Distillation for Text Reranking
AI 摘要
腾讯（Yuanbao 团队）提出 Yuanbao-Reranker，基于 Qwen3 构建 0.6B、4B、8B 三档通用稠密重排序模型。方法上结合答案优先与查询类型引导的文档接地合成、三阶段数据筛选，并用教师分数间隔采样难/中/易负例，构造 275 万条 1:3 排序组；训练时联合优化逐点交叉熵与有界最小间隔 hinge 目标，支持跨规模与同规模蒸馏及逐层自适应模型合并。在 CMTEB-R、MMTEB-R、BEIR、MLDR、MTEB-Code 五个基准上，各参数档均领先：0.6B 达 70.96 平均分超过 1B–4B 基线，4B 达 73.63 超过 8B 教师，8B 达 74.21 取得该档最优。
正文节选
Yuanbao-Reranker Abstract High-precision text reranking is a critical component in modern search, recommendation, and agentic retrieval-augmented generation (RAG) systems. We present Yuanbao-Reranker, a family of general-purpose dense rerankers spanning 0.6B, 4B, and 8B parameter scales built upon the Qwen3 backbone. To construct diverse and reliable ranking supervision at scale, we pair answer-first and query-type-guided document-grounded synthesis with a three-stage curation funnel that verifi