RTaC: A Generalized Framework for Tooling

In the rapidly evolving domain of Large Language Models (LLMs), integrating tool usage remains a formidable challenge, particularly when it comes to the dynamic selection and sequencing of tools in response to complex queries. Addressing this, we introduce Reimagining Tooling as Coding (RTaC), a groundbreaking framework that transforms tool usage into a coding paradigm. Inspired by recent advancements [18], RTaC conceptualizes tools as Python functions within a dual-agent system [2], significantly enhancing LLMs’ tool usage efficiency. Our comprehensive experiments reveal that RTaC enables coding-based LLMs, such as DeepSeek and CodeLlama, to achieve and surpass GPT-4 benchmarks in cost-effectiveness and latency without compromising on handling intricate tool sequencing with conditional and iterative logic. This research not only sets a new benchmark for tooling efficiency in LLMs but also opens new avenues for the application of LLMs in complex problem-solving scenarios, heralding a significant leap forward in the functionality and versatility of LLMs across diverse domains.