IM-RAG (Multi-Round Retrieval-Augmented Generation Through Learning Inner Monologues)

IM-RAG (Multi-Round Retrieval-Augmented Generation Through Learning Inner Monologues) is an advanced LLM-centric approach that enhances traditional RAG systems by implementing an iterative reasoning process that mimics human inner dialogue. This novel framework addresses key limitations in conventional RAG paradigms including limited flexibility in integrating diverse Information Retrieval systems, constrained interpretability during multi-round retrieval processes, and lack of end-to-end optimization. The system consists of three core components: a Reasoner (LLM) that serves as the central reasoning engine, a Retriever that gathers external information, and a Refiner that bridges gaps between the reasoning and retrieval modules. During the inner monologue process, the LLM continuously evaluates whether to query for additional information or provide a final answer based on conversational context. The entire framework is optimized through Reinforcement Learning with a Progress Tracker providing mid-step rewards, while answer prediction is separately optimized via Supervised Fine-Tuning. This architecture enables more flexible integration of various IR systems while providing enhanced interpretability through learned inner monologues that reveal the system’s reasoning process.