AtomicRAG: Atom-Entity Graphs for Retrieval-Augmented Generation

Collection of recent IR/recSys papers (arXiv 2604.*) introducing novel retrieval, recommendation, and evaluation methods. Highlights: AtomicRAG proposes Atom-Entity Graphs that store fine-grained "knowledge atoms" and use personalized PageRank + relevance filtering to improve RAG retrieval and reasoning. CaST-POI and ADS-POI propose candidate-conditioned spatiotemporal and multi-state decomposition approaches for next-POI recommendation. TASTE provides a multimodal music-recommendation dataset and MuQ-token feature-aggregation method. MATRAG presents a multi-agent, knowledge-augmented RAG for可