Visual Reasoning via Tool-Supervised RL: Curriculum Training Separates Tool Mastery from Task Solving

arXiv·medium signal

ToolsRL proposes a two-stage RL curriculum for multimodal LLMs: Stage 1 trains purely on tool-specific rewards (zoom-in, rotate, flip, draw) to master tool calling; Stage 2 adds accuracy rewards while allowing tool use. Separating tool mastery from task solving avoids optimization conflicts between heterogeneous objectives. Uses simple native visual tools whose supervision is easy to collect. From Amazon AGI.