ToolWeave: Framework for Realistic Multi-Turn Tool-Calling Dialogue Synthesis
ToolWeave is a structured framework designed to synthesize realistic multi-turn tool-calling dialogues for training large language models (LLMs) as autonomous agents. Existing synthetic data generation pipelines often produce unrealistic dialogues by chaining superficially compatible tools and generating dialogues in one shot, leading to parameter hallucination and underrepresentation of multi-step interactions. ToolWeave addresses these issues by constructing tools with built-in dependencies and filtering workflows based on alignment with user goals, thereby reducing parameter hallucination and supporting realistic multi-step workflows.
Key facts
- ToolWeave is a framework for synthesizing multi-turn tool-calling dialogues.
- It addresses unrealistic dialogues in existing synthetic data pipelines.
- Existing pipelines chain superficially compatible tools and generate dialogues in one shot.
- ToolWeave constructs tools with built-in dependencies.
- It filters workflows based on alignment with user goals.
- It reduces parameter hallucination.
- It supports realistic multi-step workflows.
- The framework is introduced in arXiv paper 2605.12521.
Entities
Institutions
- arXiv