Trace-Free+ Curriculum Learning Improves LLM Tool Use

A new framework called Trace-Free+ addresses the bottleneck of tool description quality in LLM-based agents. While most research focuses on improving the agent itself, the quality of tool interfaces—often written for humans—limits performance. Existing methods require expensive per-tool pipelines involving query synthesis, trajectory execution, annotation, and prompting. Trace-Free+ uses curriculum learning to progressively transfer supervision from trace-rich settings (where execution traces are available) to trace-free deployment, improving scalability and generalization to unseen tools without per-tool pipelines. The framework is designed to rewrite tool descriptions for reliable agent use.