vitercik-lab/DSR-Bench-spatial

DSR-Bench-spatial扩展了DSR-Bench中的3种数据结构（K-D堆、K-D树、几何图）以支持不同维度和数据分布的变体。它包含了1D、2D、3D和5D版本的所有三种数据结构，以及K-D树的3种非均匀数据分布（月亮、圆圈、水滴）版本，每个版本都包含短、中、长提示，总共450个问题。DSR-Bench-spatial旨在展示LLM在空间推理方面的优势和弱点。K-D树的非均匀数据任务也帮助识别模型在从更标准的数据模式推广到非标准数据模式方面的鲁棒性。

DSR-Bench-spatial extends 3 data structures in DSR-Bench (K-D Heap, K-D Tree, Geometric Graphs) into variants in terms of dimensionality and data distribution. It contains the 1D, 2D, 3D, and 5D data versions of all three data structures, and 3 non-uniform data distributions (moons, circles, blobs) versions of K-D Tree, all containing short, medium, and long prompts, yielding a total of 450 questions. DSR-Bench-spatial is designed to highlight LLMs’ strengths and weaknesses in terms of spatial reasoning. Non-uniform data tasks for K-D Tree also help identify the model’s robustness in terms of generalizing from more standard data patterns to less standard data patterns.