Gecko: A Simulation Environment with Stateful Feedback for Refining Agent Tool Calls.
Zhang · Li · Xing · Apostolopoulos · Lee · Zheng
A simulation environment that checks tool-call validity, synthesizes schema-adherent responses, and assesses task completion — enabling LLMs to refine tool calls without the cost or risk of real tool execution. Improves tool-calling performance across GPT-4o, GPT-5, and Gemini-3.0-pro on BFCLv3 and τ²-bench.
Read on arXiv