SAGE Benchmark for Multimodal Grounding Evaluation

SAGE reformulates existing grounding benchmarks into a unified 10-candidate image-selection protocol covering noun, verb, and entity-centered grounding. The study introduces Gloss-Augmented Reasoning (GAR), a training-free framework that uses definitions and semantic verification to improve performance. Using Gemma3-12B, GAR improved Hits@1 scores, particularly in VerSe and VWSD settings.