SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue

AI 导读

论文提出SpeakerMem-R1双轨记忆系统，通过区分个人与群体视角解决多人对话中的记忆混淆问题，实验显示其在多个基准测试中表现优于现有方法。

SpeakerMem-R1 introduces dual-track memory for multi-party dialogue, achieving state-of-the-art results on benchmark tasks by addressing attribution and relational understanding challenges.

重点速览

• 双轨记忆机制区分个人与群体视角存储对话信息 Dual-track memory separates personal and group perspectives
• 解决消息归属判定与关系理解的核心瓶颈 Addresses message attribution and relational understanding
• 在GroupMemBench等基准测试中准确率达69.2% Achieves 69.2% accuracy on SocialMemBench
• 通过实体/事件/时间维度融合多源证据 Fuses evidence across entities/events/times

一句话：双轨记忆架构显著提升了多人对话状态重建的准确性。 / Dual-track architecture significantly improves state reconstruction in multi-party dialogues.

推荐理由 方法创新性强，实验结果验证了记忆系统在复杂社交场景中的有效性。