On October 26, 2026, a research paper titled “SAGE: Sense-Aligned Multimodal Grounding Evaluation with Gloss-Augmented Reasoning” was published on the arXiv platform (arXiv ID: 2610.31838v1), officially introducing the Sense-Aligned Grounding Evaluation (SAGE) benchmark and the Gloss-Augmented Reasoning (GAR) method.
SAGE is a unified benchmark for 10 candidate image selection tasks, reorganizing the original three categories—VWSD (Visual Word Sense Disambiguation), VerSe (Verb Grounding), and WikiDiverse (Entity Diversity Grounding)—into a shared term-centric protocol. It covers multimodal grounding scenarios across three semantic centers: nouns, verbs, and entities. Its design incorporates challenging negative samples that are easily confused, as well as a context-exchange version of the VerSe diagnostic task, aiming to evaluate models’ ability to achieve deep semantic grounding of target terms, rather than relying solely on literal text–image matching.
The paper also introduces GAR—a training-free reasoning framework that combines definition grounding with constraint-guided semantic verification. Experiments on the Gemma3-12B model show that definition enhancement consistently improves performance across all three subsets; constraint-guided reasoning yields significant gains on VWSD and VerSe (especially VerSe), but only limited improvements on WikiDiverse; and the joint use of both approaches achieves the best overall macro-average performance, boosting Hits@1 from 65.90 to 71.22.
The study indicates that strong VWSD performance cannot be uniformly transferred to other grounding settings, highlighting the necessity of SAGE for unified evaluation and fine-grained attribution analysis. All conclusions are based on the experimental results reported in the paper, without incorporating external data or unspecified model variants.