Training multimodal search agents requires tasks in which visually grounded search is genuinely necessary to determine the answer. A new arXiv paper argues that synthetic questions often fail this requirement: textual identity clues and unused target evidence allow shortcuts, so an agent can answer without looking at the relevant visual information. These shortcuts weaken the training signal, the abstract states.
The paper introduces RACE, a method whose name stands for Referential Abstraction and Compositional Evidence. According to the title, the approach targets the shortcut problem by abstracting references and composing evidence across modalities, though the abstract as listed is truncated and does not detail the mechanism.
Because the source is a single preprint abstract, the claims here are limited to the problem statement and the proposed framing. No experimental results or comparisons are available in the provided text.