M-RAG: Making RAG Faster, Stronger, and More Efficient

This feed collects recent (Mar 31 2026) arXiv submissions focused on improving Retrieval-Augmented Generation (RAG) and multimodal/structured retrieval for LLMs and downstream systems. Key technical trends across papers: (1) replacing chunk-based retrieval with compact, intent-aligned key/value decompositions (M-RAG) to reduce fragmentation and improve stability; (2) introducing intermediate ‘‘abstract’’ representations organized in tree structures plus an improved Cuckoo Filter for much faster entity location (Bridge-RAG, reporting ~15.65% accuracy gain and 10x–500x retrieval speedups vs. bas