Do Multimodal Agents Really Benefit from Tool Use? A Systematic Study of Capability Gains
A new study questions the effectiveness of tool use in multimodal AI agents, suggesting that observed benchmark gains may not stem from genuine capability improvements. Researchers found that agents like Thyme and DeepEyesV2 showed minimal consistent gains from tool access, with most problems solvable even without tools. The study indicates that these agents may be learning to mimic tool-calling patterns rather than truly leveraging tools for enhanced problem-solving. AI
IMPACT Challenges the assumption that tool use inherently improves AI agent capabilities, prompting a re-evaluation of current evaluation methods.