Overview
- A cluster of arXiv papers published Thursday proposes multi‑agent and graph‑structured Retrieval‑Augmented Generation (RAG) changes to make LLM outputs more grounded and checkable.
- LegalGraphRAG builds a hierarchical legal graph and a three‑role pipeline—Researcher to retrieve evidence, Auditor to verify sources, and Adjudicator to synthesize findings—to improve transparent legal reasoning.
- MixRAGRec uses a mixture‑of‑experts retrieval agent and a knowledge‑alignment agent to route queries to the right graph granularity for recommendations, aiming to cut noise from flattening graph data into text.
- GroundedCache introduces four cheap safety gates—query similarity, retrieved‑evidence overlap, source‑version validity, and lexical or judge support—and reports dropping unsafe cached answers to near zero on HotpotQA and large reductions under document drift.
- SilentRetrieval shows a semantically preserving data‑poisoning attack can hijack RAG at very low injection rates (a sampled Wikipedia test kept 74.2% hit rate at 0.016% poisoning) and the papers report that combined retrieval and generation defenses cut attacks but add latency, so operators face trade‑offs in accuracy, safety, and speed.