OnePred: Proactive AI Anticipates User Queries in Multi-Turn Conversations

Researchers propose OnePred, a system enabling large language models to predict a user's next query in multi-turn conversations without re-reading full history. The model uses a recursively updated memory to track intent trajectory across topics, unresolved needs, and interest shifts. This approach addresses the efficiency-quality trade-off in next-query prediction, where naive concatenation of dialogue history grows token consumption linearly, while truncation loses cross-turn context. OnePred bounds per-turn cost independently of history length. The work is published on arXiv (2605.23668) and aims to advance proactive interaction in conversational AI.