The fundamental goal of agentic visual reasoning is to improve the success rate of multimodal large language models (MLLMs) on complex tasks, rather than merely equipping them with a sophisticated yet inefficient reasoning paradigm. In this work, we rethink agentic visual reasoning through two key dimensions of tool use: Mode Adaptiveness (MA) and Tool Effect (TE). Mode Adaptiveness characterizes whether an MLLM can recognize when tools are truly necessary and invoke them accordingly, thereby avoiding unnecessary computational overhead while improving performance on challenging problems that require tool assistance. Tool Effect characterizes the actual impact of tool use: tools should extend the model's capabilities on problems unsolvable through text-only reasoning, while avoiding additional errors on problems that the model can already solve without tools. We conduct a comprehensive analysis to quantify these two properties and empirically reveal that existing agentic visual reasoning models exhibit limited Mode Adaptiveness, while the gains produced by tool use on hard examples are largely offset by the harm introduced on easy examples that the models can already solve. Motivated by these observations, we propose Beacon, a novel agentic visual reasoning model that achieves stronger overall performance, improved Mode Adaptiveness, and genuine tool-induced performance gains. At the core of Beacon are the Necessity-Aware Adaptive Reward and the Hint-Guided Capability Expansion mechanism in the reinforcement learning stage, which respectively encourage adaptive tool invocation based on task necessity and strengthen the model's tool-use capability on the most challenging problems. Extensive experiments across diverse benchmarks demonstrate the strong overall performance of Beacon and its substantial improvements in both Mode Adaptiveness and Tool Effect.
Alex: Welcome to another episode of ResearchPod. Today, we're looking at a new model called Beacon, which is designed to change how AI agents use external tools.
Sam: Right. The central puzzle here is something researchers call the "tool-use paradox." AI models often reach for tools on simple questions where they don't actually help, but then fail to use them on hard questions where they're genuinely needed.
Alex: So the paper is asking why these AI systems are so poor at judging when a tool is actually worth using?
Sam: Exactly. Think of a student who uses a calculator for basic addition—wasting time—but then refuses to use it for complex calculus, where they really need it. Beacon is designed to be a kind of "smart apprentice" that learns to recognise exactly when a tool is necessary and when it isn't.
Alex: How does the team actually teach the model to make that judgement?
Sam: They use a training approach called reinforcement learning. The basic idea is a reward system: the model earns a high score for solving a simple problem on its own, without any tools. But if it struggles, it's nudged toward using a tool—and it only gets the full reward if that tool-assisted attempt actually succeeds. The paper calls this a "Necessity-Aware Adaptive Reward."
Alex: Oh—so the model is being trained to treat tool-use as a deliberate, strategic choice rather than just a default habit.
Sam: Precisely. And there's a second piece to the training. For the hardest problems, the researchers inject expert-generated hints that show the model exactly how to use a tool correctly. Then they remove those hints, so the model has to learn to replicate that reasoning on its own. They call this "Hint-Guided Capability Expansion"—essentially training wheels that get taken away once the model has learned to balance.
Alex: That's a neat approach. But did the experiments actually show it leads to better performance in practice?
Sam: Yes. They tested Beacon across thirteen different benchmarks—a wide range of tasks—and it achieved the best average performance compared to existing models. The key reason is that it keeps what the researchers call "tool harm" very low. That's when reaching for a tool on a simple question actually makes the answer worse, not better.
Alex: I want to make sure I understand the training properly. How do they stop the model from just memorising which specific questions need a tool, rather than learning the underlying logic?
Sam: That's where a third mechanism comes in, which they call "Mode-conditioned Online Labeling." Rather than giving the model a fixed set of right answers to memorise, they evaluate its performance live, during training. The model tries a problem both ways—once without a tool, once with—and the system compares the outcomes in real time. If the tool-assisted path is the only route to a correct answer, the model learns to value that path. It's being taught to prioritise accuracy over habit.
Alex: So it's not learning a rulebook. It's learning a kind of reasoning process.
Sam: That's a good way to put it. And that's the core contribution of the paper. By focusing on when to think and when to reach for a tool, the researchers argue that the decision itself—the judgement call—is just as important as the tools available. A model that chooses well will consistently outperform one that simply has access to better tools.
Alex: It turns tool-use from a messy, hit-or-miss feature into something that functions as a genuine extension of the model's own reasoning.
Sam: That's exactly the aim. And if this approach holds up across more real-world applications, it suggests a meaningful shift in how we think about building capable AI systems—not just giving them more tools, but teaching them the wisdom to use those tools well.
Alex: Thanks for listening to ResearchPod.