Bo Li, Mingda Wang, Gexiang Fang, Shikun Zhang, Wei Ye

Researchers propose GRIP, a framework that integrates retrieval decisions directly into the token generation process rather than treating retrieval as a separate external step. The model learns to emit control tokens that signal when to retrieve information, how to reformulate queries, and when to stop, all within a single autoregressive pass. Trained on structured datasets covering answerable, partially answerable, and multi-hop questions, GRIP matches or exceeds strong RAG baselines and approaches GPT-4o performance while using substantially fewer parameters.