BedBench

Evaluation data, model outputs, and probing features for the paper Where Do Vision-Language Models Fail at 3D Reasoning? Every task is posed on rendered views of 100 procedurally generated voxel objects (T1-T12) or 100 multi-object scenes (T13). Code (generator, evaluation harness, probing, fine-tuning) is released separately.

Configurations

Each configuration is one task under one visual condition (the default setting is listed first).

Task	Configurations
T1	uncolor, uncolor_axes, color
T2	uncolor, uncolor_axes
T3	uncolor_axes, color_axes
T4	color, color_axes
T5	color, color_axes
T6	uncolor, uncolor_axes
T7	color, color_axes
T8	uncolor, uncolor_axes
T9	uncolor, uncolor_axes
T10	color, color_axes
T11	invalid_uncolor, invalid_uncolor_axes, invalid_color, valid_uncolor, valid_uncolor_axes, valid_color
T12	uncolor, uncolor_axes, color
T13	invalid_uncolor, invalid_color, valid_uncolor, valid_color

T5_finetune holds the 800 training and 100 validation objects of the T5 fine-tuning experiment (generator seed 100; no object overlaps the benchmark).

from datasets import load_dataset
ds = load_dataset("wnkh/BedBench", "T5_color", split="train")

The evaluation harness reads this repository directly: python -m bed_eval run --dataset wnkh/BedBench ....

Extras

File	Contents
extras/samples.tar.gz	the 100 objects (manifest with all targets, every rendered view, projections, distractors) and the 100 T13 scenes, as written by the generator
extras/outputs.tar.gz	every model response used in the paper (12 open-weight models x 2 modes x 5 runs, 4 frontier models, image-removal, answer-format, reasoning-budget, factorized-T5, and fine-tuned runs) with ground truth;bed_eval score outputs reproduces the paper's numbers
extras/probing_features.tar	mean-pooled features of the 14 frozen encoders used in the probing analysis

License

CC BY 4.0.