BrickVision

The benchmark of "Where and Why is Grounded Reasoning Useful?" (arXiv:2611.08485, project page, code).

BrickVision tests fine-grained visual search and comparison. Each item is a photo of a procedurally generated brick scene: an isometric render restyled into a photograph with the layout unchanged, so the answer is exact (it comes from the scene geometry) and the render's segmentation masks still apply pixel for pixel. 4 tasks x 250 items:

task	the model must	answer
Crop ID (crop_id)	find which of ten look-alike structures an enlarged inset shows and point to its outlined piece	1 point
Fine-grained counting (counting)	count the 2x4 bricks with a 1x1 on top / a 1x1 at a short end / lying crosswise on their support	integer
Duplicate (duplicate)	point to the only two stacks with the same design	2 points
Unpaired (unpaired)	point to the one stack that has no twin	1 point

Evaluate a model

git clone https://github.com/elorian-ai/why-grounding.git && cd why-grounding && uv sync --all-extras
hf download elorian-ai/BrickVision --repo-type dataset --local-dir data/brickvision
python -m brickvision.evaluate score --prompts prompts.jsonl       # one {task, scene_id, image, prompt} row per item
python -m brickvision.evaluate score --predictions preds.jsonl     # rows {task, scene_id, response}; [--pixel]

The prompts ask for [{"point_2d": [x, y]}] in 0-1000 coordinates (--pixel: image pixels). response is the final answer without the reasoning, since every coordinate pair in it counts as a point. A pointing reply is correct iff it has exactly one point per target, each inside that target's mask; a counting reply iff its integer equals the gold count.

Files

folder	contents
restyled/<task>/	the benchmark: items.jsonl and samples/*.png; question is the prompt used for the Gemini models (points as [y, x]); score --prompts writes the prompts for every other model
synthetic/<task>/	the renders before restyling: items.jsonl, scenes.jsonl (scene specifications), meta.json (generator settings), samples/, and masks/ (label maps; not needed for counting)

Paths inside an items.jsonl are relative to its folder.

Results (paper Table 4, exact accuracy %)

model	Crop ID	Fine-grained counting	Duplicate	Unpaired	Overall
GPT-6 Astra	90.0	25.2	74.8	84.8	68.7
Claude Opus 5.5	40.4	16.8	19.6	27.6	26.1
GPT-5.6 Sol	35.2	8.4	11.6	20.4	18.9
Gemini 3.8 Flash	20.4	15.2	8.8	17.6	15.5
Claude Opus 5	26.8	10.4	8.4	11.6	14.3
Qwen3.8-Max	6.4	2.0	13.2	12.4	8.5
Qwen3.5-9B (base)	1.5	8.5	2.4	8.4	5.2
Finetuned Direct	8.1	16.0	8.3	7.7	10.0
Finetuned Text Reasoning	85.2	66.4	39.1	54.1	61.2
Finetuned Grounded Reasoning	91.6	95.3	64.7	79.9	82.9

The fine-tuned models are Qwen3.5-9B; their checkpoints and every model output are in elorian-ai/why-grounding. Models trained on data from brickvision.generate (like the fine-tuned rows above) should be reported separately from models that were not.

How it was made

The scenes come from python -m brickvision.generate (seeds 66002, 61002, 64004 and 56002 for Crop ID, counting, Duplicate and Unpaired). Each render was restyled into a photograph with gemini-3.1-flash-image-preview, and a paired judge (gemini-3.7-flash) checked that the layout was preserved: up to 3 attempts over 4 style templates; no item was dropped.

Licence: CC BY 4.0.

@article{yan2026grounded,
  title   = {Where and Why is Grounded Reasoning Useful?},
  author  = {Yan, David and Xue, Le and Huang, Forrest and Zhang, Qiuyi and Tran, Dustin and Dai, Andrew M. and Yang, Yinfei},
  journal = {arXiv preprint arXiv:2611.08485},
  year    = {2026}
}