Overview

• SciTreeRAG and SciGraphRAG use a hierarchical corpus structure and an LLM-built knowledge graph to preserve cross-document context, with proof-of-concept on CERN’s LHCb literature.
• Legal and regulatory QA tests a One-SHOT chunk selection strategy under a token budget and an iterative Reasoning Agentic retrieval loop, adding fixes for query drift and retrieval laziness.
• VaccineRAG introduces a Chain-of-Thought dataset and evaluation that requires per-sample reasoning and trains with Partial-GRPO to strengthen discrimination against harmful or irrelevant retrievals.
• A CFA-based study evaluates 1,560 official mock exam questions across Levels I–III and reports that reasoning-focused models lead in zero-shot and that a curriculum-grounded RAG pipeline boosts accuracy on complex items.
• Across papers, retrieval precision remains the main failure point, and authors say they will release code and datasets to enable replication and real-world adoption.