Beacon: Agentic Visual Reasoning Model Improves Tool Adaptiveness and Effectiveness

KlingTeam · hf · 2026-07-31

Beacon is a novel agentic visual reasoning model addressing Mode Adaptiveness (MA) and Tool Effect (TE). Existing models show limited MA, and tool gains on hard examples are offset by harm on easy ones. Beacon uses Necessity-Aware Adaptive Reward and Hint-Guided Capability Expansion in RL to encourage adaptive tool invocation and strengthen tool-use capability. Experiments show stronger overall performance and substantial improvements in MA and TE.

Original post →

More from Multimodal

Multimodal channel →