LIBRA: Programmatic Induction of Executable Tool Libraries
Story 4, from arXiv
LIBRA is a new framework that transforms flat tool catalogs into executable libraries by dynamically composing and testing tool interactions, improving retrieval performance across multiple benchmarks.
Key points
- LIBRA dynamically constructs executable tool libraries from flat catalogs without dependency annotations or task-specific training.
- It uses LLMs to propose tool contracts and tests compositions in a sandbox to retain valid runtime relations.
- LIBRA improves one-shot Recall and NDCG by 7.9 and 3.7 percentage points on AppWorld, BFCL, and SciToolEval.
- It supports both one-shot and agentic retrieval, with gains of 4.5 and 2.9 points after ten retrieval iterations on ToolReAGt.
- The framework requires no user input during library construction, making it efficient for pre-query setup.
Why it matters
As AI agents increasingly rely on tool catalogs, LIBRA offers a more effective way to organize and retrieve tools, reducing the need for manual annotation and improving task performance in complex scenarios.
Counterpoint
LIBRA's effectiveness depends on the quality of the initial tool catalog and the LLM's ability to accurately propose valid tool compositions.
Sources
This summary was written by a language model and may contain errors. Read the original for the full story.