ARTFEED — Contemporary Art Intelligence

CoRAG: Reranker and Generator as Peer Decision-Makers

ai-technology · 2026-04-27

A new framework called Cooperative Retrieval-Augmented Generation (CoRAG) redefines the relationship between reranker and generator in RAG systems. Instead of the traditional asymmetric dependency where generation quality hinges on reranking results, CoRAG treats both components as peer decision-makers jointly optimized toward a shared task objective. This cooperative approach encourages the reranker and generator to work in concert, improving document reranking and generation for better final responses. Experimental results show good generalization and improved generation stability.

Key facts

  • CoRAG treats reranker and generator as peer decision-makers
  • Traditional RAG uses a ranking-centric, asymmetric dependency paradigm
  • CoRAG jointly optimizes reranker and generator behaviors
  • Experimental results show good generalization and improved generation stability
  • The framework is proposed to overcome limitations of existing RAG systems

Entities

Institutions

  • arXiv

Sources