SpeakerMem-R1 Tops EverMemBench at 62.33% with Dual-Track Multi-Party Dialogue Memory
zju · hf · 2026-09-24
SpeakerMem-R1 targets two bottlenecks in multi-party long-term dialogue memory: speaker attribution/relational understanding and state reconstruction from interleaved histories. It stores speaker-labeled verbatim messages plus derived states in dual tracks organized as person- and group-level views, merging evidence by entity, event, and time at query time. Writer-R1, trained with SpeakerLevenshtein and speaker-conditioned GRPO, reduces attribution errors while enabling local deployment.
- 62.33% on the public EverMemBench leaderboard, best reported among latest SOTA frameworks; 47.9% on GroupMemBench, 69.2% on SocialMemBench
- 70.85% across all 1,986 LoMo questions as a two-person boundary test
- RL lifts SFT Writer mean accuracy from 57.38% to 68.20% on a 305-question controlled eval
- Ablations show verbatim/structured tracks and person/group views are complementary
More from Research
- Debate clarifies: is LLM 'model-directedness' just a 'current speaker' representation? — rgblong · 2026-09-24
- IEEE launches ASLI 2027 speech and language intelligence conference in Singapore — shinjiw_at_cmu · 2026-09-24
- Researchers dissect Anthropic emotion concepts paper: separate self/other speaker representations — rgblong · 2026-09-24
- IEEE launches ASLI 2027, first conference on audio, speech and language intelligence, in Singapore — shinjiw_at_cmu · 2026-09-24
- Bioinformatician resurfaces 21-year-old "antedisciplinary science" essay, says a new field is being born — lpachter · 2026-09-24
- Heng Li's minisplice: a 7k-parameter CNN halves minimap2 splice junction error rate — anshulkundaje · 2026-09-24