quchangle1/LLM-Tool-Survey

What it solves

The field of tool learning with large language models (LLMs) is growing fast, but the research is scattered. This project gathers the key papers and organizes them according to a systematic survey, making it easier for newcomers to understand the landscape and for researchers to find relevant work.

How it works

The project is built around a survey paper (available on arXiv) that classifies tool learning into two main questions: why tool learning is beneficial and how tool learning is implemented. The "why" covers the benefits of using tools (like search engines, databases, math solvers, Python interpreters) and the benefits of the learning paradigm itself (interpretability, robustness). The "how" breaks the workflow into four stages: task planning, tool selection, tool calling, and response generation. The README lists hundreds of papers, each tagged with its venue and a link to the arXiv entry, and organizes them under these categories. It also includes sections on benchmarks, evaluation methods, challenges, and other resources.

Who it's for

Researchers, students, and industrial developers who want a structured overview of tool learning with LLMs, or who need a convenient index of papers to reference or build upon.

Highlights

• Comprehensive taxonomy: papers are grouped by why tool learning helps and how it is done, with four workflow stages.
• Extensive paper list: dozens of entries across categories like knowledge acquisition, expertise enhancement, automation, interaction, and benchmarks.
• Direct links to every paper on arXiv, with venues and years listed.
• Includes the full survey paper, accepted at Frontiers of Computer Science, and a Chinese summary for broader access.
• Open to contributions: issues and pull requests are welcome to keep the list current.