检索片段训练实现高效 QMSum 查询聚焦会议摘要
原标题：Retrieved-Span Training for Efficient Query-Focused Meeting Summarization on QMSum
AI 摘要
该论文针对 QMSum 查询聚焦会议摘要任务，指出该基准缺少统一评分器导致结果难以比较，因此在一个实现下重新评分或生成 15 个系统并发布逐查询预测。核心发现是训练与推理输入对齐很关键：一个 406M Fusion-in-Decoder 专用模型从长输入切换到 2000 词检索片段后损失 6.30 ROUGE-1，但在该片段机制上微调后可恢复，测试得分 36.33 ROUGE-1，与 1.2B 系统 35.41 无统计显著差异，且参数约为三分之一、峰值推理内存不到一半。论文还报告在统一提示和评分器下，该 406M 模型超过五个专有托管模型至少 6.2 ROUGE-1，并发布协议、评分器、预测和训练产物。
正文节选
Retrieved-Span Training for Efficient Query-Focused Meeting Summarization on QMSumThanks: Agentic coding tools supported implementation, experiment execution, and prose drafting under the author’s direction. The author remains responsible for all content; a fuller disclosure appears before the references. Abstract Query-focused meeting summarization requires answering from transcripts that exceed small-model context windows. QMSum provides no scorer, making reported results difficult to compare.