SAFE-QE：过滤并抑制LLM查询扩展中的幻觉词
原标题：Tracing and Taming Hallucinated Terms via Subsequence-Association Filtering and Breadth-Aware Selection for LLM Query Expansion
AI 摘要
该论文提出 SAFE-QE，一种无需训练、与检索器无关的查询扩展方法，利用子序列关联理论识别并抑制 LLM 生成的幻觉扩展词。方法通过语料条件化的跨上下文可复现性打分，对在随机查询上下文中反复出现但缺乏伪相关集支持的词降权，并以约束覆盖最大化选择近似广泛且一致的词集。在 TREC 和 BEIR 基准上，SAFE-QE 的 nDCG@10 与生成式和语料引导扩展无显著差异，但每查询注入的不支持词仅 4.3 个，远低于 Query2doc 的 27.8 和 HyDE 的 57.5。
正文节选
Tracing and Taming Hallucinated Terms via Subsequence-Association Filtering and Breadth-Aware Selection for LLM Query Expansion Abstract Large language models (LLMs) are now a standard source of query expansion terms but they hallucinate. They inject non existent entities and off-topic terms that add ranking noise and can hurt strong retrievers. Existing remedies such as corpus grounding, mutual verification and preference alignment treat hallucinated terms as an undifferentiated nuisance to be