Research reveals that multi-step tool-use reinforcement learning in large language models often collapses, resulting in abrupt performance drops and failed tool-invocation structures. This catastrophic collapse occurs due to the instability of reinforcement learning alone, which can lead to limited gains in complex tasks. To address this issue, supervisory signals have been introduced, demonstrating the ability to fix the collapse and enhance model capabilities. The use of supervisory signals provides a more stable and effective approach to reinforcement learning, allowing large language models to perform complex tasks with improved reliability. The findings have significant implications for the development of large language models, as reinforcement learning can reshape both capability and risk surfaces, with security implications following the hype cycle1. This matters to practitioners because the instability of reinforcement learning can have severe consequences, making it crucial to develop more robust and reliable methods for enhancing model capabilities.