Rigorous 3D spatial reasoning benchmark testing autonomous multi-modal AI agents inside Portal physics puzzles.
Freely accessible open-source 3D spatial reasoning benchmark dataset and evaluation harnesses.
Portal2Bench addresses the glaring shortcoming of contemporary AI benchmarks that evaluate models solely on flat static text puzzles rather than authentic three-dimensional physical causal reasoning. The acclaimed video game Portal requires actors to navigate rigorous Newtonian dynamics, momentum preservation, and spatial trajectories across multi-step puzzle chambers. Portal2Bench codifies these environments into an uncompromising agent benchmark, rigorously testing whether vision-language models can infer spatial depth, project momentum acceleration trajectories, and orchestrate complex physical action sequences.