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

TL;DR

The SpeakerMem-R1 model addresses bottlenecks in multi-party dialogue by storing speaker-labeled messages and derived states in a dual-track memory. This approach improves message attribution and relational understanding. The model achieves state-of-the-art results on several benchmarks.

Why It Matters

Accurate multi-party dialogue understanding has significant implications for applications like chatbots, virtual assistants, and social media analysis. Improved performance in this area can enhance user experience and facilitate more effective human-computer interaction. This research contributes to the development of more sophisticated conversational AI systems.

Key Findings

• SpeakerMem-R1 outperforms existing models on GroupMemBench, SocialMemBench, and EverMemBench
• The dual-track memory approach improves message attribution and relational understanding
• The model achieves a binary accuracy of 62.33% on the EverMemBench leaderboard