File size: 4,800 Bytes
340e0ec
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
411f594
340e0ec
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
---
license: cc-by-4.0
task_categories:
- visual-question-answering
language:
- en
pretty_name: GameBind counterfactual twins
size_categories:
- 10K<n<100K
---

# GameBind counterfactual twins

Rendered two-agent scenes and their counterfactual twins, released with the anonymous submission "Counterfactual
Twins Identify Perception and Binding Errors in Vision-Language Models".

Each base scene shows a target agent and a competitor agent in an 8x8 tile world. Each agent performs one action
(pick up, push, hit or go to) on one object. Every action leaves a visible end state: a picked-up object is carried,
a pushed object slides and leaves skid marks, a hit object cracks and leaves debris, and go to changes nothing. A twin
re-renders the same scene with one controlled change, and all other pixels stay the same. Comparing a model's answers
across the twins of a scene separates two causes of a wrong answer: missing the target's action and guessing, or
seeing it and attributing the competitor's action to the target.

## Files

| file | contents | clips | questions | images |
|---|---|---|---|---|
| `gamebind_twins_full.zip` | 1,000 base scenes and 300 near-condition scenes with their main twins | 5,362 | 16,086 | 42,896 |
| `gamebind_twins_ext.zip` | further twins of the same 1,000 scenes | 2,318 | 6,954 | 18,544 |
| `gamebind_twins.zip` | the API subset: a separate draw of 300 base scenes and 150 near-condition scenes | 1,717 | 5,151 | 13,736 |
| `gamebind_twins_ext_api.zip` | further twins of the API subset | 677 | 2,031 | 5,416 |
| `colorshape_twins.zip` | colour-shape displays, 600 base scenes (easy display) | 4,200 | 4,200 | 3,601 |
| `colorshape_twins_hard.zip` | colour-shape displays, 600 base scenes (hard display) | 4,200 | 4,200 | 3,601 |
| `per_scene/` | every model's answers per scene and twin, from which all results of the paper are computed | | | |

Unzip all files in one folder. Each zip creates `data/<set>/` with:

- `clips.json`: one entry per clip, with the base scene, the twin, the events, the agents and the paths of its frames;
- `queries.json`: the questions asked on each clip, with their options and the correct answer;
- `checks.json`: the pixel checks of every twin;
- `summary.json`: the generator settings and the check summary;
- `frames/`: the images.

The frame paths in `clips.json` are relative to the folder where you unzip. A GameBind clip has 8 frames
(`f0.png` to `f7.png`, 512x512 pixels), and the models in the paper saw the last frame, `f7.png`. A colour-shape clip
is one 512x512 image.

## Twins

| twin | change |
|---|---|
| `S` | base scene |
| `minus_c` | the competitor stays idle at its start tile |
| `minus_cue` | the target moves as in S, but its action leaves no trace |
| `minus_cue_minus_c` | both changes |
| `minus_ccue` | only the competitor's trace is deleted, the competitor still moves |
| `degraded_cue` | the target's trace is weakened |
| `swap` | the two agents exchange their events |
| `placebo` | an irrelevant change to the floor pixels |
| `S_repro` | a fresh rendering of S, used to check that rendering is exact |

In the colour-shape displays, the twins grey out the competitor shape (`minus_c`), the target shape (`minus_cue`) or
both, add a grey shape elsewhere (`placebo`), or exchange the two colours (`swap`). A blind grey image (`blind`) is
also included.

## Questions

On every GameBind clip three questions are asked, each with four options. The joint question asks what the target
agent did, and its options are the true event, two answers that mix the target's and the competitor's action and
object, and the competitor's whole event. The object question and the action question ask for one part. Options are
shown under rotated orders, and all twins of a base scene share the same orders. The colour-shape question asks for
the colour of the target shape, with the target's colour, the competitor's colour and two absent colours as options.

## Model answers

`per_scene/` holds one gzip CSV per model and scene set: 19 open models on all scenes, and GPT-4.1, GPT-4o, GPT-5.5
and seven open models on the API subset. Each row is one question about one twin of one base scene, with the answer
probability of every option. The subfolders `colorshape/` and `photo_twins/` hold
the answers on the colour-shape displays and on the real-photo twins (numbers only). `per_scene/README.md` describes
the columns.

## Notes

- All images are rendered. The dataset contains no photographs and no personal data.
- The renderer regenerates every frame bit for bit. Code: https://anonymous.4open.science/r/counterfactual-twins-578F
- The API subset is a separate draw from the 1,000 scenes, although the scene names repeat. A scene is identified by
  its set and its `base_id`.

## Licence

CC BY 4.0.