Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Researchers introduced Beacon, a multimodal large language model that adaptively decides when tool use is needed for visual reasoning.

Prior models often overused tools or saw limited gains because improvements on hard cases were offset by errors on easy ones.

Beacon uses reinforcement learning to encourage necessity-based tool invocation and improve performance on difficult problems.

The approach aims to make multimodal systems more accurate, efficient, and reliable by using tools only when they truly help.