RefSpatial-Bench

🎉 RefSpatial-Expand-Bench is officially released! The new version not only extends indoor scenes (e.g., factories, stores), but also introduces brand-new outdoor scenarios (e.g., streets, parking lots) — enabling more comprehensive evaluation of spatial referring tasks. 👉 Try it now: RefSpatial-Expand-Bench 🏆 The paper associated with this benchmark, RoboRefer , has been accepted to NeurIPS 2025 ! Thank you all for your attention and support! 🙌 RefSpatial-Bench: A Benchmark for Multi-step Spatial Referring         Welcome to **RefSpatial-Bench**, a challenging benchmark based on real-world cluttered scenes to evaluate more complex multi-step spatial referring with reasoning. 🎯 Task Split Location Task: This task contains **100** samples, which requires model to predicts a 2D point indicating the **unique target object**. Placement Task: This task contains **100** samples, which requires model to predicts a 2D point within the **desired free space**. Unseen Set: This set comprises **77** samples from the Location/Placement task, specifically designed to **evaluate model generalization after SFT/RFT training on RefSpatial**, as it includes novel spatial relation combinations not present in RefSpatial. ⚠️ Warning: If your model is not trained with RefSpatial, Unseen set should not be used for evaluation. 🧠 Reasoning Steps We introduce *reasoning steps* (`step`) for each benchmark sample as the number of anchor objects and their spatial relations that help constrain the search space. A higher `step` value reflects greater reasoning complexity and a stronger need for spatial understanding and reasoning. 📁 Dataset Structure We provide two formats: Hugging Face Datasets Format `data/` folder contains HF-compatible splits: `location` `placement` `unseen` Each sample includes: | Field | Description | | :------- | :----------------------------------------------------------- | | `id` | Unique integer ID | | `object` | Natural language description of target