NMR-Bench

A visual reasoning benchmark for Visual Scaffolding. It tests whether a model can keep track of object identity and spatial continuity while reasoning over an image.

📄 Paper: Visual Scaffolding: Toward Native Multimodal Reasoning

Task	Challenge	Test images
Burden Counting	Count and compare objects in crowded scenes	420
Path Tracing	Follow curves through crossings to match endpoints	360

Task-specific checkpoints: Burden Counting · Path Tracing.

🖼️ A look at the benchmark

Two hard examples from the test set are shown below. The queries and ground truths match the released annotations.

Burden Counting	Path Tracing

Burden Counting · counting_test_T2c_1000040

Query: How many butterflies in this image are partially covered by another object? Give your final answer as a single integer inside \boxed{}, for example \boxed{7}.

GT: There are \boxed{21} covered butterflies.

Path Tracing · path_test_cell060_seed00060000

Query: The image contains multiple curves, each connecting two endpoints in the image. The endpoints in this image are labeled: {blue_circle, cat, crown, diamond_shape, dog, frog, hammer, heart, light_bulb, mushroom, star, sun}. Each label appears exactly once. Your task is to identify all connection pairs.

Trace each curve carefully. Pay attention to crossings where lines overlap — at each crossing, the correct continuation should maintain smooth direction and curvature consistency. After completing all traces, verify your results to ensure they are consistent and correct.

Output the connection pairs in the format: \boxed{(label_a, label_b), (label_c, label_d), ...}

GT: Based on the traces above, the connection pairs are: \boxed{(blue_circle, star), (cat, diamond_shape), (crown, heart), (dog, light_bulb), (frog, mushroom), (hammer, sun)}.

🚀 Get started

With Python 3.12, download the test sets and run the scorer from the GitHub repository:

git clone https://github.com/GeoX-Lab/VisualScaffolding.git
cd VisualScaffolding
pip install -e .
pip install huggingface_hub
hf download ShaoRun/nmr-bench --repo-type dataset --local-dir benchmark
python -m visual_scaffolding score \
  --task burden_counting \
  --benchmark benchmark \
  --predictions predictions/counting.jsonl \
  --output outputs/counting-summary.json

The scorer expects a JSONL prediction for every test sample, with sample_id and response fields. Use --task path_tracing for the other task. See the GitHub README for model inference and the full evaluation workflow.

📂 Data and license

Each task has an images/ directory, metadata.jsonl for the Hub's image viewer, and test.jsonl for the companion evaluator. Records contain the image path, question, answer, difficulty, and sample checksum. Use the answer only for scoring; pass the image and question to the model. Field definitions are in SCHEMA.md, and file checksums are in manifest.json.

Compositions and annotations are licensed under CC BY-SA 4.0. The images include OpenMoji and Twemoji artwork under their respective licenses. See DATA_LICENSE.md and NOTICE.md for attribution.