6 papers
Knowledge Graphs
Graph-structured memory and retrieval for agents that need to connect facts.
- Papers
- 6
- Read time
- 1.7h
- Avg length
- 17m
Graph-Augmented Large Language Model Agents: Current Progress and Future Prospects
When your LLM agents start hallucinating plans, forgetting context, fumbling tools, and stepping on each other, the fix is almost always the same primitive you already half-use — a graph — and this survey is the field map for where to put it.
- Read
- 16 min
- Source
- arXiv:2507.21407
Graph-R1: Towards Agentic GraphRAG Framework via End-to-end Reinforcement Learning
Instead of retrieving a graph once and hoping the LLM can reason over it, Graph-R1 turns retrieval into a multi-turn "think → query → read → rethink → answer" agent loop and trains that whole loop end-to-end with reinforcement learning, so a small 7B model beats GPT-4o-mini GraphRAG pipelines on multi-hop QA.
- Read
- 18 min
- Source
- arXiv:2507.21892
GRIL: Knowledge Graph Retrieval-Integrated Learning with Large Language Models
GRIL trains a graph retriever and an LLM *together* so the retriever learns to fetch the exact facts the LLM actually needs to answer a question, instead of guessing what looks relevant and hoping it helps.
- Read
- 18 min
- Source
- arXiv:2509.16502
Knowledge Graph Engineering for Multi-Agentic Systems: The Anthropic Playbook
Replace the entire trained-model NLP pipeline (NER + relation classifier + resolution heuristics) with four Claude prompts that turn a pile of documents into a queryable, provenance-carrying knowledge graph — and use that graph as the shared memory, fact-checker, and persistent world model your multi-agent systems have been missing.
- Read
- 20 min
RAG vs. GraphRAG: A Systematic Evaluation and Key Insights
A head-to-head benchmark proving RAG and GraphRAG aren't rivals but complements — plain RAG wins on detailed/single-hop questions, GraphRAG wins on multi-hop reasoning and diverse summaries — so the real win is a router or a merge, not picking one.
- Read
- 16 min
- Source
- arXiv:2502.11371
ReaGAN: Node-as-Agent-Reasoning Graph Agentic Network
Turn every node in a graph into its own little LLM agent that decides — per node, per layer — whether to gather info from its graph neighbors, retrieve semantically similar nodes from anywhere in the graph, predict, or do nothing, and you can match trained Graph Neural Networks using a frozen LLM with zero fine-tuning.
- Read
- 16 min
- Source
- arXiv:2508.00429