---
dataset_info:
- config_name: real
  features:
  - name: question_index
    dtype: int64
  - name: image_id
    dtype: string
  - name: question
    dtype: string
  - name: answer
    dtype: string
  - name: question_category
    dtype: string
  - name: question_type
    dtype: string
  - name: target_object
    dtype: string
  - name: objects
    sequence: string
  - name: id
    dtype: string
  - name: image
    dtype: image
  splits:
  - name: test
    num_bytes: 1383586783.875
    num_examples: 5439
  download_size: 559763548
  dataset_size: 1383586783.875
- config_name: sim
  features:
  - name: question_index
    dtype: int64
  - name: image_id
    dtype: string
  - name: question
    dtype: string
  - name: answer
    dtype: string
  - name: question_category
    dtype: string
  - name: question_type
    dtype: string
  - name: target_object
    dtype: string
  - name: objects
    sequence: string
  - name: id
    dtype: string
  - name: image
    dtype: image
  splits:
  - name: test
    num_bytes: 10278779977.424
    num_examples: 22644
  download_size: 2894555522
  dataset_size: 10278779977.424
configs:
- config_name: real
  data_files:
  - split: test
    path: real/test-*
- config_name: sim
  data_files:
  - split: test
    path: sim/test-*
---

# PercepTax — Benchmarking Physical Intelligence through Cross-Property Reasoning in Vision-Language Models

**PercepTax** is an **open-ended** visual-question-answering benchmark that tests whether
vision-language models can reason *across* physical object properties — material, shape,
function, affordance — together with spatial relations and compositional / counterfactual
reasoning. In each question the object of interest is marked by a **colored box**, and the
model must answer in free form (no multiple-choice options in the prompt).

![PercepTax samples](assets/preview.png)

<p align="center">
  <a href="https://perceptual-taxonomy.github.io"><img alt="Project Page" src="https://img.shields.io/badge/Project_Page-2563EB?style=for-the-badge&logo=googlechrome&logoColor=white"></a>
  <a href="https://arxiv.org/abs/2511.19526"><img alt="Paper" src="https://img.shields.io/badge/arXiv-2511.19526-B31B1B?style=for-the-badge&logo=arxiv&logoColor=white"></a>
  <a href="https://github.com/XingruiWang/PercepTaxBench"><img alt="Code" src="https://img.shields.io/badge/Code-181717?style=for-the-badge&logo=github&logoColor=white"></a>
  <a href="https://huggingface.co/datasets/TaxonomyProject/SimulationMetadata"><img alt="Sim Metadata" src="https://img.shields.io/badge/Sim_Metadata-FFD21E?style=for-the-badge&logo=huggingface&logoColor=black"></a>
</p>

## Subsets

| Config | Source | Questions | Images |
|--------|--------|-----------|--------|
| `real` | Real photos (OpenImages) with 3D annotation | 5,439 | 2,516 |
| `sim`  | Rendered simulated indoor scenes | 22,644 | 14,499 |

Each subset has a single `test` split. Question categories:
`taxonomy_reasoning` · `taxonomy_description` · `spatial_relation`
(fine types in `question_type`, e.g. `compositional_set_subtraction_container`,
`repurposing_shield_concept`, `spatial_above_below`).

## Fields

| Field | Type | Description |
|-------|------|-------------|
| `image` | Image | scene image; the queried object is marked by a colored box |
| `question` | string | open-ended question |
| `answer` | string | ground-truth answer (object name, or a spatial direction) |
| `question_category` | string | coarse category |
| `question_type` | string | fine-grained type |
| `target_object` | string | object the question is about |
| `objects` | list[string] | objects referenced in the scene |
| `image_id` | string | source image / scene id |
| `id` / `question_index` | string / int | identifiers |

## Quick start

```python
from datasets import load_dataset

ds = load_dataset("RyanWW/PercepTaxBench", "real", split="test")   # or "sim"
ex = ds[0]
ex["image"]                      # PIL.Image (object marked by a colored box)
print(ex["question"], "->", ex["answer"], f"({ex['question_category']})")
```

## Running a model (inference + scoring)

Open-ended answers are scored by **normalized exact match** with an optional **LLM judge**
for paraphrases. Minimal, model-agnostic loop:

```python
from datasets import load_dataset

ds = load_dataset("RyanWW/PercepTaxBench", "real", split="test")

def predict(image, question):
    # plug in any VLM (HF transformers, or an API such as GPT-4o / Gemini / Claude)
    ...

def norm(s): return " ".join(s.lower().split())

correct = 0
for ex in ds:
    pred = predict(ex["image"], ex["question"])
    correct += norm(pred) == norm(ex["answer"])      # swap in an LLM judge for paraphrase credit
print("accuracy:", correct / len(ds))
```

For a full, reproducible evaluation harness (per-category accuracy, LLM-judge scoring across
many VLMs), use the **VLMEvalKit integration** in the code repo:
👉 https://github.com/XingruiWang/PercepTaxBench#inference--evaluation

## License & citation

[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/).

```bibtex
@inproceedings{lee2026perceptax,
  title     = {PercepTax: Benchmarking Physical Intelligence through Cross-Property Reasoning in Vision-Language Models},
  author    = {Lee, Jonathan and Wang, Xingrui and Peng, Jiawei and Ye, Luoxin and Zheng, Zehan and Zhang, Tiezheng and Wang, Tao and Ma, Wufei and Chen, Siyi and Chou, Yu-Cheng and Kaushik, Prakhar and Yuille, Alan},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}
```