Reproduce “Affordable AI Assistants with Knowledge Graph of Thoughts” on OpenResearch

About this paper

Maciej Besta, Lorenzo Paleari, Jia Hao Andrea Jiang, Robert Gerstenberger, You Wu, Jón Gunnar Hannesson, Patrick Iff, Ales Kubicek and 10 more

Researchers at ETH Zurich developed KGoT, an architecture integrating Large Language Model reasoning with dynamically constructed knowledge graphs, to create more affordable and effective AI assistants. The system significantly reduces operational costs by over 36x while enabling smaller LLMs to achieve superior performance on complex tasks, including outperforming larger models and state-of-the-art baselines on benchmarks like GAIA and SimpleQA.

Key insights

• Externalizing LLM reasoning into an explicit, evolving Knowledge Graph significantly enhances interpretability and auditability while improving reasoning capabilities for complex, multi-step tasks.
• Structuring task-relevant knowledge within a KG allows smaller, more affordable LLMs to achieve performance levels comparable to, or even exceeding, much larger and more expensive models.
• Combining a modular architecture with asynchronous execution, graph parallelism, and distributed processing provides a robust and scalable solution for real-world AI assistant deployment.

Summary by alphaXiv.