To overcome the Limitations of the Retrieval-Augmented Generation (RAG) model, we propose the Collaborative Chain-of-Agents (CoCoA) framework, which aims to enhance the synergy between the model's inherent knowledge and externally retrieved knowledge. CoCoA builds on the multi-agent RAG framework, CoCoA-zero, and performs conditional knowledge induction and answer inference. CoCoA utilizes a long-chain training strategy that fine-tunes the LLM by synthesizing multi-agent inference trajectories extended from CoCoA-zero. This allows the model to explicitly integrate and jointly leverage inherent and retrieved knowledge.
