SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue
The paper claims stronger multi-party memory by keeping both verbatim speaker-labeled messages and structured person/group state.
SpeakerMem-R1 is built to track who said what, who it concerns, shared group information, and changes over time. Its query-time retrieval combines evidence by entity, event, and time across those two memory tracks. The authors report 47.9% on GroupMemBench, 69.2% on SocialMemBench, and 61.9% on EverMemBench, plus 62.33% on EverMind-AI’s public EverMemBench leaderboard. They also say reinforcement learning raised a structured memory writer’s mean accuracy from 57.38% to 68.20% in a 305-question controlled evaluation. HF Daily Papers' note
SpeakerMem-R1 is built to track who said what, who it concerns, shared group information, and changes over time. Its query-time retrieval combines evidence by entity, event, and time across those two memory tracks. The authors report 47.9% on GroupMemBench, 69.2% on SocialMemBench, and 61.9% on EverMemBench, plus 62.33% on EverMind-AI’s public EverMemBench leaderboard. They also say reinforcement learning raised a structured memory writer’s mean accuracy from 57.38% to 68.20% in a 305-question controlled evaluation. HF Daily Papers' note
score 4