Research map / Multi-agent systems and games / LLM agents
GraphPlanner: Graph-Based Agentic Routing for LLMs
Tao Feng, Haozhen Zhang, Zijie Lei, Peixuan Han, Jiaxuan You
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GraphPlanner, a graph-based agentic router that decides which LLM or agent to call at each step of multi-agent workflows, extending LLM routing to agentic settings.
Problem
LLM routing has mostly targeted single-turn queries; agentic settings requiring task planning, multi-round cooperation among heterogeneous agents, and memory use lack suitable routing methods.
Approach
Models routing for agentic workflows as a heterogeneous graph over queries, agents and workflow state, so a learned policy can select agents at each step while balancing performance and cost. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use agentic workflow router GraphPlanner instead of best baseline router | LLM routing · Phase 2 average accuracy | accuracy: 54.3% -> 63.6% | Improved |
| use GraphPlanner instead of Router-R1 on cost | LLM routing · Phase 2 average cost | cost: 76.3 -> 605.0 | Hurt |
| use GraphPlanner to pick backbones inside fixed workflows | Phase 1 workflows · Phase 1 Depth=1 Width=3 average | accuracy: 54.80% -> 58.60% | Improved |
| generate query-specific workflows (Phase 2) instead of optimizing within fixed workflows (Phase 1) | GraphPlanner · average accuracy | accuracy: 58.6 -> 63.6 | Improved |
| use GraphPlanner on commonsense reasoning (Phase 2) | LLM routing · Phase 2 CS | accuracy: 82.7% -> 78.0% | No clear effect |
| add historical memory to the router (ablation) | GraphPlanner · Phase 2 five scenarios | accuracy | Improved |
| use heterogeneous GARNet instead of homogeneous GNN (ablation) | GraphPlanner · Phase 2 five scenarios | accuracy | Improved |
| use transductive inference (reuse stored training interactions) instead of inductive (ablation) | GraphPlanner · Phase 2 five scenarios | accuracy | Improved |
| add Thinker and Verifier roles with training (ablation) | GraphPlanner · Phase 2 Math | accuracy: 67.0% -> 70.5% | Improved |
| offer unseen roles at test time with no training (ablation) | GraphPlanner · Phase 2 Math | accuracy: 67.0% -> 68.5% | Improved |
| apply GraphPlanner to unseen tasks zero-shot | GraphPlanner · LogicGrid, MGSM, CommonGen | accuracy: 58% -> 78% | Improved |
Closest papers by meaning
- Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.90)
- DiSRouter: Distributed Self-Routing for LLM Selections (ICLR 2026 · similarity 0.89)
- Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- GraphFlow: A Graph-Based Workflow Management for Efficient LLM-Agent Serving (ICML 2026 · similarity 0.88)
- Learning Query-Aware Budget-Tier Routing for Runtime Agent Memory (ICML 2026 · similarity 0.88)
- Benchmarking Agentic Workflow Generation (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
