Simulating Complex Multi-Turn Tool Calling Interactions in Stateless Execution Environments

Researchers have introduced DiGiT-TC, a novel data generation method designed to simulate complex multi-turn tool calling interactions within stateless execution environments. While synthetic data is crucial for tuning cost-effective language models, previous frameworks typically assumed the availability of stateful environments to validate interaction outcomes. This assumption limits applicability in real-world scenarios, such as enterprise settings with strict data security requirements or systems integrating tools from diverse sources. DiGiT-TC addresses this gap by employing a unique generation pattern that implicitly represents specific tool calls within user requests, effectively mimicking the characteristics of conversations generated through stateful search processes. The study validates this approach using standard tool calling benchmarks, demonstrating significant performance improvements even in traditional stateful problem settings. Published on arXiv, this work offers a valuable resource for enhancing the capabilities of smaller language models in handling intricate tool-use tasks without relying on persistent state maintenance, thereby broadening the scope for secure and flexible AI deployment in various industrial applications.