Hi authors,
Thank you for sharing this insightful work! I'm particularly interested in the trajectory synthesis process described in Section 3.3, where models (e.g., GPT-4o, Gemini) generate step-by-step reasoning traces with tool-invocation markers.
To better understand the implementation, could you clarify:
- The exact prompts used to elicit tool-invoking trajectories from models (e.g., few-shot examples or templates)?
- Whether the code for synthesizing and executing these trajectories is available in the repository? If not, would you consider releasing it?
Looking forward to your feedback, and thanks again for the great work!