Language Agents as Optimizable Graphs

arxiv:2402.16823

Original source · Paper JSON

Agents, open-endedness, AGI · 2024-02-26

Cites (2) · Cited by (5)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4

Contents

• GPTSwarm: Language Agents as Optimizable Graphs
• Abstract
• Keywords:
• 1 Introduction
• 2 GPTSwarm
• 2.1 Language Agents as Graphs
• 2.2 Graph Definition
• Single language agent as a graph.
• Swarm of language agents as a composite graph.
• 2.3 Edge Optimization
• 2.3.1 Problem Reformulation
• 2.3.2 Solution Parameterization
• 2.3.3 Optimization Algorithm
• 2.4 Node Optimization
• 2.5 General Applicability
• 3 Experiments
• 3.1 MMLU
• 3.1.1 Adversarial setting
• 3.1.2 Collaborative setting
• 3.2 Mini Crosswords
• 3.3 HumanEval
• 3.4 GAIA
• 4 Related Work
• 4.1 LLM-based Autonomous Agents
• 4.2 Language Agents with Graphs
• 4.3 Optimizing LLM Inference and Self-Improvement
• 5 Conclusion
• Impact Statement
• Author Contributions
• Acknowledgements
• Appendix A The GPTSwarm Framework
• A.1 The Vision
• A.2 Class Diagram
• Appendix B Swarm examples
• Appendix C More Visualizations
• Appendix D Additional Experiments
• D.1 Comparing Our Method with Multiagent Debate and DYLAN on MMLU
• D.2 Applying Node Optimization after Edge Optimization
• Appendix E Experimental Details
• E.1 MMLU
• E.1.1 Hyper-parameters & Prompts
• E.1.2 Adversarial swarm optimization case study
• E.1.3 Swarm Roles
• E.2 Mini Crosswords
• E.2.1 Agents Setting
• E.2.2 Hyper-parameters & Prompts
• E.3 HumanEval
• E.3.1 The Node Optimization Method
• E.3.2 Hyper-parameters & Prompts
• E.4 GAIA
• E.4.1 Agent Setting
• E.4.2 Hyper-parameters & Prompts
• Appendix F Resource Requirements
• Appendix G Limitation and Future Work
• Instructions for reporting errors