README.md

---
license: mit
language:
- en
pretty_name: VRE-Bench (Visual Rule Exploration)
size_categories:
- n<1K
task_categories:
- image-text-to-text
- reinforcement-learning
tags:
- benchmark
- vision-language-models
- interactive-reasoning
- rule-induction
- exploration
- human-baseline
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: id
    dtype: int64
  - name: exploration_image
    dtype: image
  - name: exploration_image_description
    dtype: string
  - name: test_image
    dtype: image
  - name: test_image_description
    dtype: string
  - name: layout
    dtype: string
  - name: control
    dtype: string
  - name: reward_source
    dtype: string
  - name: hidden_info
    list: string
  - name: theme
    dtype: string
  - name: mechanics
    list: string
  - name: hidden_rule
    dtype: string
  - name: human_explorations
    list:
    - name: annotator_id
      dtype: int64
    - name: attempts
      list:
      - name: actions
        list: string
      - name: rewards
        list: int64
      - name: infos
        list: string
  - name: human_difficulty
    dtype: int64
  splits:
  - name: test
    num_bytes: 14717879
    num_examples: 100
  download_size: 14576569
  dataset_size: 14717879
---

# VRE-Bench: Visual Rule Exploration

100 hand-designed visual rule exploration puzzles that each have a hidden rule that an agent has to discover
through experiments. Each puzzle consists of two phases: 
1. In an **exploration phase**, the agent sees an `exploration_image`,
has to submit action sequences, and gets back integer points per action as a reward.
2. In a **test phase**, it has to apply
what it has learned to a `test_image` of the same puzzle.

VRE-Bench measures how efficiently vision-language models learn hidden rules, at which step they fail, and whether they
know when to stop exploring.

A test answer is scored by min-max normalizing its total points with the puzzle's exact reward
bounds. The action sets, reward logic, bounds and the evaluation harness live in
[vreBench](https://github.com/flowun/vreBench).

## Fields

| Field | Type | Description |
|---|---|---|
| `id` | int | Puzzle id (1–100), the key into vreBench's puzzle implementations |
| `exploration_image`, `test_image` | image | The two puzzle states |
| `exploration_image_description`, `test_image_description` | string | Python-literal textual descriptions of the images (grids, lists, dicts), complete enough to solve the puzzle without the image |
| `hidden_rule` | string | Natural-language statement of the hidden rule. Is not shown to the model. |
| `layout` | string | `graph` (40) · `geometric` (26) · `grid` (19) · `stateful` (15) |
| `control` | string | `avatar` (55) · `manipulation` (30) · `selection` (15) |
| `reward_source` | string | `collect` (42) · `condition` (39) · `differential` (19) |
| `hidden_info` | list[string] | Rule parameters the image does not show: `none`, or any of `values`, `pairings`, `remapping` |
| `theme` | string | Depicted scenario (e.g. `islands-bridges`, `mine`), or `abstract` |
| `mechanics` | list[string] | Optional extra tags (e.g. `rotation`, `irreversible`) |
| `human_difficulty` | int | 1–5 from human performance: the mean normalized test score over the five checkpoints, ranked and cut into five bins of 20 puzzles each |
| `human_explorations` | list | Per annotator: `annotator_id` and `attempts`, the 10 recorded exploration attempts (`actions`, `rewards`, `infos`). Currently one annotator, on all 100 puzzles |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("flowun/vre", split="test")
```

To run a model:

```bash
pip install git+https://github.com/flowun/vreBench.git
vre --model <model-name>   # any OpenAI-compatible endpoint
```

## Citation

If you use VRE-Bench, please cite:

```bibtex
@thesis{TODO,
  title  = {TODO: paper title},
  author = {TODO},
  year   = {TODO},
  url    = {TODO}
}
```