id
stringlengths
5
5
image
imagewidth (px)
1.92k
1.92k
category
stringlengths
11
22
question
stringlengths
1.11k
1.53k
choices
listlengths
4
4
answer
stringclasses
4 values
q0002
identify_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
D
q0006
identify_rightmost
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. <8>", "B. <1>", "C. <12>", "D. <9>" ]
D
q0007
relative_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
C
q0008
order_leftmost
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. <6>, <10>, <5>, <8>", "B. <6>, <5>, <10>, <8>", "C. <10>, <5>, <8>, <6>", "D. <10>, <5>, <6>, <8>" ]
D
q0013
pick_closer
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. <13>", "B. both are equidistant", "C. <10>", "D. cannot be determined" ]
C
q0014
order_closest
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. <1>, <5>, <13>, <3>", "B. <3>, <13>, <5>, <1>", "C. <1>, <3>, <5>, <13>", "D. <1>, <13>, <5>, <3>" ]
C
q0015
identify_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
C
q0017
identify_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. forward — same direction as ego (12 o'clock)", "B. leftward — perpendicular left (9 o'clock)", "C. rightward — perpendicular right (3 o'clock)", "D. backward — toward ego (6 o'clock)" ]
A
q0018
pick_closer
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. <2>", "B. both are equidistant", "C. <6>", "D. cannot be determined" ]
A
q0019
relative_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
B
q0020
identify_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
C
q0021
identify_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. forward — same direction as ego (12 o'clock)", "B. rightward — perpendicular right (3 o'clock)", "C. leftward — perpendicular left (9 o'clock)", "D. backward — toward ego (6 o'clock)" ]
D
q0022
relative_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
C
q0024
identify_type
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. car", "B. pedestrian", "C. suv", "D. light_truck" ]
C
q0025
identify_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
B
q0028
relative_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. cannot be determined", "B. perpendicular to each other", "C. roughly opposite directions", "D. roughly the same direction" ]
C
q0029
identify_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
B
q0031
relative_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. roughly the same direction", "B. perpendicular to each other", "C. cannot be determined", "D. roughly opposite directions" ]
D
q0033
identify_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
B
q0034
identify_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. backward — toward ego (6 o'clock)", "B. leftward — perpendicular left (9 o'clock)", "C. forward — same direction as ego (12 o'clock)", "D. rightward — perpendicular right (3 o'clock)" ]
B
q0038
identify_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. leftward — perpendicular left (9 o'clock)", "B. forward — same direction as ego (12 o'clock)", "C. rightward — perpendicular right (3 o'clock)", "D. backward — toward ego (6 o'clock)" ]
A
q0040
order_closest
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. <22>, <11>, <8>, <9>", "B. <11>, <8>, <9>, <22>", "C. <11>, <9>, <22>, <8>", "D. <8>, <9>, <11>, <22>" ]
D
q0044
pick_closer
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. both are equidistant", "B. <2>", "C. cannot be determined", "D. <3>" ]
B
q0045
relative_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
D
q0048
identify_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. leftward — perpendicular left (9 o'clock)", "B. backward — toward ego (6 o'clock)", "C. rightward — perpendicular right (3 o'clock)", "D. forward — same direction as ego (12 o'clock)" ]
D
q0049
relative_distance
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Very close (0-2m)", "B. Close (2-10m)", "C. Medium (10-30m)", "D. Far (30m+)" ]
C
q0050
relative_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. roughly opposite directions", "B. roughly the same direction", "C. cannot be determined", "D. perpendicular to each other" ]
A
q0052
embodied_collision
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Yes", "B. No", "C. cannot be determined", "D. only if the object moves" ]
A
q0053
identify_frontmost
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. <12>", "B. <6>", "C. <14>", "D. <8>" ]
C
q0055
relative_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
D
q0056
relative_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. perpendicular to each other", "B. roughly the same direction", "C. cannot be determined", "D. roughly opposite directions" ]
D
q0060
identify_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. forward — same direction as ego (12 o'clock)", "B. backward — toward ego (6 o'clock)", "C. leftward — perpendicular left (9 o'clock)", "D. rightward — perpendicular right (3 o'clock)" ]
D
q0066
relative_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
C
q0069
order_leftmost
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. <17>, <3>, <27>, <9>", "B. <9>, <17>, <27>, <3>", "C. <9>, <3>, <27>, <17>", "D. <17>, <9>, <27>, <3>" ]
D
q0070
identify_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
D
q0072
identify_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. leftward — perpendicular left (9 o'clock)", "B. forward — same direction as ego (12 o'clock)", "C. backward — toward ego (6 o'clock)", "D. rightward — perpendicular right (3 o'clock)" ]
B
q0073
relative_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
D
q0075
identify_closest
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. <2>", "B. <4>", "C. <1>", "D. <3>" ]
C
q0080
relative_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. cannot be determined", "B. perpendicular to each other", "C. roughly the same direction", "D. roughly opposite directions" ]
D
q0081
order_closest
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. <2>, <1>, <3>, <4>", "B. <4>, <1>, <2>, <3>", "C. <3>, <4>, <1>, <2>", "D. <1>, <2>, <3>, <4>" ]
D
q0083
relative_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
A
q0084
identify_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. forward — same direction as ego (12 o'clock)", "B. leftward — perpendicular left (9 o'clock)", "C. backward — toward ego (6 o'clock)", "D. rightward — perpendicular right (3 o'clock)" ]
D
q0086
identify_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
D
q0087
identify_frontmost
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. <1>", "B. <4>", "C. <3>", "D. <6>" ]
D
q0088
identify_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
B
q0094
identify_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
C
q0095
identify_position
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. nearby", "B. ahead-left", "C. behind-right", "D. ahead" ]
B
q0096
relative_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
C
q0098
identify_distance_long
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. Close (0-20m)", "B. Medium (20-50m)", "C. Far (50-80m)", "D. Very far (80m+)" ]
B
q0100
identify_heading
You are answering a 3D spatial-reasoning question from a SINGLE monocular driving image with numbered bounding boxes. Frontier models routinely get these right by luck while reasoning incorrectly: they lean on flat-image shortcuts ('lower in the frame = closer', 'bigger box = nearer', 'left in the image = to my left') ...
[ "A. backward — toward ego (6 o'clock)", "B. rightward — perpendicular right (3 o'clock)", "C. forward — same direction as ego (12 o'clock)", "D. leftward — perpendicular left (9 o'clock)" ]
C