In the rapidly evolving field of artificial intelligence, multimodal large language models (MLLMs) are increasingly expected to tackle complex visual reasoning tasks. However, a fundamental challenge remains: knowing when to use external tools and how to use them effectively. A new research paper introduces Beacon, a model designed to address this challenge by focusing on two critical dimensions: Mode Adaptiveness (MA) and Tool Effect (TE).
Mode Adaptiveness measures whether a model can recognize when tools are truly necessary, avoiding unnecessary computational overhead while improving performance on challenging problems. Tool Effect, on the other hand, evaluates the actual impact of tool use—tools should extend capabilities on unsolvable problems without introducing errors on simpler ones the model can already handle.
The researchers conducted a comprehensive analysis revealing that existing agentic visual reasoning models often lack Mode Adaptiveness. They found that the performance gains from tool use on hard examples are frequently offset by the harm caused on easy examples, where the model already performs well without tools.
To overcome these limitations, Beacon introduces two novel mechanisms during reinforcement learning: the Necessity-Aware Adaptive Reward and the Hint-Guided Capability Expansion. These mechanisms encourage adaptive tool invocation based on task necessity and strengthen the model's tool-use capability on the most challenging problems, respectively.
Extensive experiments across diverse benchmarks demonstrate that Beacon achieves stronger overall performance, improved Mode Adaptiveness, and genuine tool-induced performance gains, marking a significant step forward in agentic visual reasoning.