OnePred: Next-Query Prediction via Recursive Intent Memory in Multi-Turn Conversations

Researchers introduced OnePred, a next-query prediction model that uses compact recursive memory instead of full dialogue history.

It is trained with a two-stage reinforcement learning pipeline to improve prediction quality in multi-turn conversations.

The team also released NQP-Bench, a new benchmark for evaluating next-query prediction.

Experiments show stronger performance and up to 22x lower per-turn token usage than full-history baselines.