无状态执行环境下复杂多轮工具调用交互模拟

Synthetic data has proven to be a valuable resource for fine-tuning smaller, cost-effective language models to handle the complexities of multi-turn tool calling conversations. While numerous frameworks and systems for producing synthetic multi-turn tool calling data have been proposed, prior works frequently assume that any tool calling interactions will take place in an execution environment that maintains state. However, many real-world applications require tool calls in stateless environments, such as constrained cloud functions, edge devices, or one-shot API invocations. In these stateless settings, each tool call is independent, without retaining any context or state from previous calls. This presents unique challenges for simulating complex interactions, as traditional simulation methods often rely on state propagation to construct coherent multi-turn dialogues. To address this, novel simulation strategies are required that can explicitly encode conversational history and contextual information into each tool call request, or design tools capable of inferring necessary state from the input/output of each invocation. Simulating multi-turn tool calling interactions in stateless environments necessitates a more refined design of tool interfaces, enabling them to process each request independently and to guide subsequent calls through explicit parameter passing or returned results. Concurrently, when generating synthetic data, it is crucial to ensure that the simulated dialogue flow possesses sufficient self-contained information at each call to support model decisions and tool execution. This approach is essential for training models to effectively utilize tools in practical stateless deployments, significantly enhancing their adaptability and robustness in resource-constrained or distributed environments.