- https://arxiv.org/pdf/2310.16421v1.pdf
- Source: arxiv
TL;DR
Graph Agent uses a multi-agent framework where a planner LLM decomposes complex graph questions into subgraph retrieval and reasoning sub-steps handled by specialized agents, outperforming few-shot GPT-4 on knowledge graph QA.
Summary
Complex knowledge graph QA requires multi-hop traversal that monolithic LLM prompting cannot reliably handle. Graph Agent addresses this with a multi-agent framework: a planner LLM decomposes questions into subgraph retrieval and local reasoning sub-steps, while specialized sub-agents handle node/edge fetching independently from reasoning. The pipeline produces explicit intermediate reasoning chains, making it interpretable. On WebQSP and CWQ, it outperforms few-shot GPT-4 and matches fine-tuned baselines using frozen models without graph-specific fine-tuning.
Key contributions
- Proposes a multi-agent decomposition framework where a planner LLM separates graph retrieval from reasoning into distinct specialized agents.
- Demonstrates that frozen LLMs with structured graph interfaces can match fine-tuned baselines on knowledge graph QA.
- Shows that explicit intermediate reasoning steps enable interpretable inspection and correction of graph traversal decisions.
- Establishes that the planner-executor pattern from code agents transfers effectively to knowledge graph reasoning.
When to cite
- When building a multi-agent system for knowledge graph QA that requires multi-hop traversal.
- When arguing that separating graph retrieval from reasoning improves precision and reduces context length.
- When comparing frozen LLM + graph interface approaches against fine-tuned graph QA models.
- When justifying explicit intermediate steps for interpretability in graph-based reasoning pipelines.