Beacon: Knowing When and How to Perform Agentic Visual Reasoning

TL;DR - Beacon is an agentic visual-reasoning model trained to invoke tools only when necessary and to use them effectively on difficult multimodal tasks. It aims to improve accuracy without the overhead and errors caused by indiscriminate tool use.

• Defines Mode Adaptiveness to measure whether a model recognizes when tools are needed.
• Defines Tool Effect to assess whether tools help hard examples without harming easy ones.
• Finds existing models’ gains on hard tasks are often offset by tool-induced errors on solvable tasks.
• Uses necessity-aware rewards and hint-guided capability expansion during reinforcement learning to improve selective, effective tool use.