个性化搜索中LLM重排序的语义与行为信号鲁棒融合
原标题：Robust Fusion of Semantic and Behavioural Signals for LLM Reranking in Personalised Search
AI 摘要
该论文研究个性化搜索中LLM交叉编码器重排序时行为信号（如点击率）的融合问题。作者发现，将查询切片统计（QSS）直接注入提示虽在特征可用时提升13.3%排序质量，但会导致模型过度依赖该特征，在稀疏或缺失场景下性能下降。为此提出双样本特征丢弃训练策略，在保留排序增益的同时将QSS移除场景性能相对提升4.0%，在线测试中搜索成功率提升约2%。
正文节选
Robust Fusion of Semantic and Behavioural Signals for LLM Reranking in Personalised Search Abstract. Personalised search sits at the boundary between retrieval and recommendation: it must satisfy explicit query intent while incorporating user context and collective interaction evidence. LLM-based cross-encoders offer a single reranking interface for these signals. In industrial search, however, relevance also depends heavily on behavioural signals such as click-through rates and historical searc