Multi-turn reinforcement learning has a silent bug most teams miss

Original: Most people training agentic LLMs with RL right now have a silently broken training loop and have no idea.

Source: x.com ↗