AGViveiros/LanteRn-Tetris-Data

LantErn Tetris Analogy Dataset

A controlled, synthetic visual-analogy benchmark used in What's Holding Back Latent Visual Reasoning? to diagnose latent visual reasoning (LVR) in vision-language models, isolated from the noise of natural-image benchmarks.

Code: GuilhermeViveiros/LantErn — synthetic/Tetris/ Model trained on this data: AGViveiros/LanteRn-3B-Tetris

Task

The model is shown a visual analogy over polyomino ("Tetris-like") pieces:

Image (A) is to image (B) as image (C) is to which of the following options? The transformation from (A) to (B) is: {rotation}.

A and B are shown as standalone floating pieces; C is shown on a grid alongside four option panels (a/b/c/d). The model must apply the same rotation observed on A→B to C and pick the matching option. Correct-answer position is balanced 25% per option.

Each sample also ships an intermediate rotation strip — A rotated to B in two 90° steps, rendered in the piece's own color — used as the ground-truth "latent" crop for LVR supervision (see bboxs / reasoning_traces.intermediate_img_path).

Example

image (eval[0]):

intermediate_image (the A→B rotation strip, ground-truth LVR crop):

{
  "sample_id": 0,
  "question": "Image (A) is to image (B) as image (C) is to which of the following options?\nThe transformation from (A) to (B) is: 180° clockwise rotation.\nOptions:\n(a) Option a\n(b) Option b\n(c) Option c\n(d) Option d",
  "answer": "a",
  "transform_type": "rotation",
  "transform_description": "180° clockwise rotation",
  "shape_A_name": "U5",
  "shape_C_name": "T4",
  "shape_A_family": "pentomino",
  "shape_C_family": "tetromino",
  "reasoning_traces": {
    "pre_visual_text_think": "By examining the A→B pair, I identify a 180° clockwise rotation. Applying this mentally to the query object,",
    "post_visual_text_think": ["The shape here is consistent with rotating the query object 180° clockwise as shown in A→B.", "", "", ""],
    "text_think": "The 180° clockwise rotation seen in A→B, when applied to the query, gives option (a).",
    "answer": "a"
  }
}

Color intermediates

Intermediate strips in this release are rendered in the shape's own color. This is the exact data AGViveiros/LanteRn-3B-Tetris was trained on — use --color_intermediate if regenerating with create_dataset.py to match this release.

Generation

python -m synthetic.Tetris.create_dataset \
    --output_dir /path/to/output \
    --seed 42 \
    --color_intermediate

Configs (shape_A, starting rotation, transform, shape_C, starting rotation) are enumerated exhaustively, shuffled with the fixed seed, and split 95/5 into train/eval with zero overlap. 46 fully-asymmetric shapes (tetromino/pentomino/ hexomino/heptomino/octomino families) are used so every A→B pair has a unique, well-defined answer.

Splits

Split	Samples
train	7,935
eval	417

(A held-out split with five unseen hexomino shapes — used for OOD generalization evals — is generated separately via --held_out and is not included here.)

Fields

Field	Type	Description
sample_id	int	Unique sample index
image	Image	Composite image: A, B (floating), C on grid + 4 options
intermediate_image	Image	The A→B rotation strip (ground-truth LVR crop)
question	str	Analogy prompt with the transform description
answer	str	Correct option letter (a–d)
bboxs	list[[x1,y1,x2,y2]]	Bounding boxes over the 4 option panels
reasoning_traces	dict	intermediate_img_path plus text reasoning fields (pre_visual_text_think, post_visual_text_think, text_think, answer)
transform_type / transform_description	str	e.g. rotation, "180° clockwise rotation"
shape_A_name / shape_C_name	str	Shape identifiers
shape_A_family / shape_C_family	str	tetromino / pentomino / hexomino / heptomino / octomino

Citation

@article{lantern_holding_back_2026,
  title={What's Holding Back Latent Visual Reasoning?},
  author={Viveiros, Guilherme and others},
  journal={arXiv preprint arXiv:2605.18445},
  year={2026}
}