SABRE-Prior

SABRE-Prior is the world-prior stress-test benchmark introduced in SABRE: Scalable and Automated Benchmarking of VLMs under Stress. It evaluates whether vision-language models follow visible evidence when that evidence conflicts with learned expectations about familiar objects, materials, scenes, and language.

Project page: https://zesearch.github.io/vlm-SABRE/
Code: https://github.com/Zesearch/vlm-SABRE
Paper: arXiv link coming soon

Dataset composition

Split	Images	Questions	Primary metric
Context	200	400	Strict four-probe pair accuracy
Texture	200	400	Strict four-probe pair accuracy
Attribute	100	100	Exact normalized count accuracy
Language	100	100	Exact multiple-choice accuracy
Total	600	1,000	Four-split macro average

The benchmark has evaluation splits only and is not intended as training data.

Loading

from datasets import load_dataset

context = load_dataset("Zesearch/SABRE-Prior", "context", split="test")
texture = load_dataset("Zesearch/SABRE-Prior", "texture", split="test")
attribute = load_dataset("Zesearch/SABRE-Prior", "attribute", split="test")
language = load_dataset("Zesearch/SABRE-Prior", "language", split="test")

Each row contains the following common fields:

• id: globally unique question identifier;
• image: image decoded by Hugging Face from file_name;
• subset: context, texture, attribute, or language;
• question and answer;
• eval_type: the official scoring contract.

Context and Texture additionally contain pair_id and probe. Language contains the four options and answer_text. Texture exposes the object and normal/counterfactual surface labels used to construct the probe.

Evaluation

Save one prediction file per split:

predictions/
├── context.jsonl
├── texture.jsonl
├── attribute.jsonl
└── language.jsonl

Every line must contain the question ID and the raw model response:

{"id": "context_001__base_source", "prediction": "yes"}

Run the included deterministic evaluator:

python evaluate.py --predictions predictions --output metrics.json

The official metrics are:

1. Context: a pair is correct only when base_source, base_target, edited_source, and edited_target are all correct.
2. Texture: a pair is correct only when base_normal, base_counterfactual, edited_normal, and edited_counterfactual are all correct.
3. Attribute: the normalized predicted count must exactly match the answer.
4. Language: the extracted A/B/C/D option must exactly match the answer.
5. Macro accuracy: the unweighted mean of the four split accuracies.

Question-level and probe-level accuracies for the paired splits are diagnostics, not the main reported Context or Texture scores.

Main results

Accuracy (%) reported in the paper:

Model	Context	Texture	Attribute	Language	Macro avg.
Claude 4.6	10	40	17	58	31.3
Kimi-k2.6	7	52	17	17	23.3
Qwen 3.5	3	46	14	29	23.0
Gemini 3.5	0	52	26	11	22.3
GPT-5.4	1	28	20	23	18.0
Grok-4.3	4	28	16	23	17.8

The mean macro-average accuracy across the six evaluated models is 22.6%.

Intended use and limitations

SABRE-Prior is intended for non-commercial research and education involving VLM evaluation, stress testing, robustness analysis, and benchmark methodology. It is not intended for model training, commercial deployment, or claims about general intelligence or real-world safety based on these scores alone.

The benchmark is intentionally adversarial and generated or edited. It does not represent the natural frequency of objects, materials, scenes, attributes, or language cues in the world. Performance should be interpreted only under the evaluation protocol above.

Content notice

The Attribute split includes synthetic images of animals and objects with non-canonical component counts. Some animal images contain unusual numbers of limbs and may be visually unsettling to some viewers. The content is provided solely for benchmark research; it does not depict real animal harm.

License and disclaimer

The dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0). The SABRE software repository has its own license.

The dataset is provided "as is," without warranties of any kind. To the extent permitted by law, the authors are not liable for claims or damages arising from its use. Users are responsible for complying with applicable law, institutional policies, the dataset license, and responsible research practice.

Citation

The arXiv identifier will be added after release.

@article{lan2026sabre,
  title   = {SABRE: Scalable and Automated Benchmarking of VLMs under Stress},
  author  = {Lan, Zixuan and Sun, Luzhe and Walter, Matthew R. and Zhou, Jiawei},
  journal = {arXiv preprint},
  year    = {2026}
}