CacheGraft：用潜空间文档增强提升 RAG
原标题：CacheGraft: Enhancing Retrieval-Augmented Generation via Latent-Space Document Enrichment
AI 摘要
该论文提出 CacheGraft 框架，将原本用于多智能体实时通信的 KV 缓存翻译技术迁移到检索增强生成（RAG）场景。方法是在离线索引阶段，用 4B 的 Qwen3-4B-Base 作为 sharer 模型、0.6B 的 Qwen3-0.6B 作为 receiver 模型，通过 C2C 融合器把大模型的 KV 缓存融合进小模型的缓存空间，查询时只需部署小模型即可注入该缓存。在五个知识密集型 QA 基准和两个本地评审模型上，CacheGraft 在 30 组对比中有 29 组优于纯文本基线，其中 17 组统计显著，且融合器未在检索数据上训练，属于零样本增益。作者认为潜空间文档增强是算力受限场景下提升 RAG 的有前景方向。
正文节选
CacheGraft: Enhancing Retrieval-Augmented Generation via Latent-Space Document Enrichment Abstract. Recent work in machine learning has demonstrated that transformer-based language model agents can communicate state not only through text, but also directly in latent space via key-value (KV) caches. In this paper, we argue that such cache-to-cache translation can provide utility beyond the original motivating application of real-time agent communication and can instead be utilized for retrieval-