Figure 1: Object semantics for open-ended naming and spatial prediction. (a) Spatial prediction produces object queries for LLM-based naming. (b) UniOSE enriches these queries with image-derived regional semantics and uses them for training-time detection matching and mask conditioning. Semantic guidance originates from the enriched query before LLM-based naming. (c) Comparison of box recall@0.75 and naming accuracy. Swin and SAM denote Open-Det Swin-T and its fully fine-tuned SAM2-T backbone variant, respectively. The protocol, metric definitions, and source results are provided in Appendix A.1.