Reproduce “GAM-RAG: Gain-Adaptive Memory for Evolving Retrieval in Retrieval-Augmented Generation” on OpenResearch

About this paper

Yifan Wang, Mingxuan Jiang, Zhihao Sun, Yixin Cao, Yicun Liu, Keyang Chen, Guangnan Ye, Hongfeng Chai

Researchers from Fudan University developed GAM-RAG, a training-free framework that enables Retrieval-Augmented Generation (RAG) systems to dynamically adapt their retrieval memory based on inference-time feedback. This approach yielded an average performance improvement of 3.95% over strong baselines, which further increased to 8.19% with 5-turn memorization, and achieved a 58% reduction in retrieval time for recurring queries.

Key insights

• Dynamic, gain-adaptive memory allows RAG systems to learn continuously from inference-time feedback, enabling persistent improvement and reducing the need for expensive offline retraining.
• Modulating memory updates with an uncertainty-aware Kalman gain provides robustness against noisy feedback and enables stable long-term adaptation by distinguishing reliable from unreliable evidence.
• A lightweight, relation-free hierarchical graph structure, combined with explicit temporal and task-oriented memory, offers an efficient and powerful foundation for multi-hop retrieval without complex knowledge graph construction.

Summary by alphaXiv.