SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue
The paper’s system keeps two kinds of memory: speaker-labeled verbatim messages and structured person/group states, then reconciles them by entity, event, and time.
SpeakerMem-R1 is aimed at multi-party conversations where memory systems need to track who said what, who it concerns, and what is shared across the group. The authors train a local Writer-R1 component to reduce attribution and update errors during memory construction. Reported results include 47.9% on GroupMemBench, 69.2% on SocialMemBench, 61.9% on EverMemBench, and 70.85% across 1,986 LoCoMo questions. Ablations say the verbatim and structured tracks, plus person-level and group-level views, each add value. ArXiv · AI/CL/LG's note
SpeakerMem-R1 is aimed at multi-party conversations where memory systems need to track who said what, who it concerns, and what is shared across the group. The authors train a local Writer-R1 component to reduce attribution and update errors during memory construction. Reported results include 47.9% on GroupMemBench, 69.2% on SocialMemBench, 61.9% on EverMemBench, and 70.85% across 1,986 LoCoMo questions. Ablations say the verbatim and structured tracks, plus person-level and group-level views, each add value. ArXiv · AI/CL/LG's note
score 4