Overview

• A Sept. 8 arXiv study contrasts a token‑budgeted One‑SHOT chunk selector with an iterative Reasoning Agentic RAG that issues and refines queries on government and legal corpora, flagging query drift and retrieval laziness.
• VaccineRAG introduces a Chain‑of‑Thought‑based benchmark and training that requires per‑sample analysis to reject misleading evidence, adds Partial‑GRPO for learning long reasoning sequences, and the authors say code and data will be released.
• SciTreeRAG builds hierarchical contexts from physics papers and SciGraphRAG retrieves via a knowledge graph to preserve cross‑document relations, demonstrated as a proof of concept on the LHCb corpus at CERN.
• A confidence‑based approach reads LLM hidden states to estimate post‑retrieval usefulness, fine‑tunes an NQ_Rerank reranker, and triggers retrieval only when needed, reporting higher accuracy for context screening and lower cost.
• A CFA‑based evaluation shows a tailored RAG pipeline with hierarchical organization and structured query generation improves financial reasoning over zero‑shot baselines across 1,560 official mock‑exam questions.