Researchers have introduced When2Speak, a new dataset and generation pipeline designed to improve how large language models handle turn-taking in multi-party conversations. The dataset contains over 215,000 examples from 16,000 conversations, focusing on the decision of when to speak versus remain silent. Supervised fine-tuning on this data significantly boosts model performance, but further improvements are achieved using reinforcement learning with asymmetric reward shaping, which reduces missed interventions and increases recall. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Enables LLMs to participate more naturally in multi-party conversations by learning appropriate turn-taking.
RANK_REASON This is a research paper introducing a new dataset and methodology for LLMs. [lever_c_demoted from research: ic=1 ai=1.0]