Reproduce “Bridging Tool Dependencies and Domain Knowledge: A Graph-Based Framework for In-Context Planning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shengjie Liu, Li Dong, Zhenyu Zhang

Amazon researchers developed a graph-based framework to automatically generate high-quality in-context planning exemplars for large language models by explicitly modeling tool dependencies and integrating domain-specific knowledge. The approach achieves high accuracy in identifying tool dependencies and improves plan generation by 9 percentage points with a Personalized PageRank algorithm.

Key insights

  • Explicitly modeling dependencies between domain-specific tools, where the output of one serves as input to another, is crucial for effective multi-step planning by LLMs.
  • Integrating structural tool dependencies with broader domain knowledge, such as usage instructions and SOPs, provides richer context, enabling LLMs to understand how and when to use tools in sequence.
  • A dense-sparse retrieval approach, combining embedding search with graph algorithms like Personalized PageRank, can effectively surface both semantically similar and implicitly connected information, overcoming limitations of pure embedding-based retrieval for complex planning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.