Beacon Targets the Overlooked Failure Mode in Agentic Visual Reasoning: Tools That Hurt on Easy Problems

arXiv·low signalXBlueskyLinkedInCopy link

The authors reframe agentic visual reasoning around two dimensions of tool use — Mode Adaptiveness, whether a multimodal LLM recognizes when tools are genuinely necessary, and Tool Effect, whether invoking tools extends capability without degrading already-solvable problems. Their finding is that existing agentic models show limited Mode Adaptiveness, so gains on hard problems get offset by newly introduced errors on easy ones, which a single aggregate score conceals. Beacon addresses this with a Necessity-Aware Adaptive Reward and Hint-Guided Capability Expansion during RL. Specific benchmark numbers are not stated in the abstract, so treat the magnitude of improvement as unverified.