Researchers have proposed a new framework called SAFE-G for answering visual questions that require knowledge from external sources. The framework uses a combination of visual and textual modalities to locate relevant documents and then employs graph retrieval to pinpoint precise evidence within those documents. A reinforcement learning strategy is used to ensure the model's reasoning process remains faithful to the retrieved evidence, assigning credit only when the selected evidence is correct. This approach has been shown to improve accuracy in knowledge-based visual question answering tasks by 8.9% and 3.5% compared to previous methods. --- Why it matters: This matters because it addresses a common challenge in AI research: ensuring that models reason accurately and faithfully based on available evidence, rather than making assumptions or generating incorrect answers. Source: https://arxiv.org/abs/2608.21796