Type: url
Status: generated
Language: en
Summary
An ICLR 2025 paper presenting ToolACE, an automatic agentic pipeline that synthesizes accurate, complex and diverse function-calling training data through tool self-evolution synthesis of a 26,507-API pool, self-guided multi-agent dialog generation calibrated by the target model’s own loss, and dual-layer rule- and model-based verification. A LLaMA-3.1-8B-Instruct model fine-tuned on the data (ToolACE-8B) is reported to reach state-of-the-art results among open models on BFCL and API-Bank, comparable to GPT-4 models, with ablations on accuracy, complexity and diversity.