RefSpatial-Bench: Results as Published

RefSpatial-Bench tests spatial referring for robotics: the model must mark one point in an image that matches a written location or placement instruction. It has 200 real-world images, 100 for Location and 100 for Placement, each with a precise mask, and more than 70 percent of the samples need multi-step reasoning, up to five steps. A predicted point succeeds when it falls inside the ground-truth mask. The reported result is the average success rate.

Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.