AI Dataset2026
Evaluates whether models can reconstruct executable 3D worlds from visual or textual inputs by providing Blender-based tasks, assets, and automated evaluation for code-driven realizations.
Answers spatial choice, scalar, and point questions by classifying among discretized candidates using pretrained attention projections from a Qwen3.5-4B backbone instead of autoregressive text generation. Trains on converted spatial QA data, adopts coarse-to-fine refinement and joint multi-stage supervision, and reports higher accuracy with several-fold faster inference than autoregressive baselines.