---
license: apache-2.0
language:
- en
tags:
- vlm
- reasoning
- multimodal
- nli
size_categories:
- n<1K
task_categories:
- visual-question-answering
---

# **NL-Eye Benchmark**

Will a Visual Language Model (VLM)-based bot warn us about slipping if it detects a wet floor? 
Recent VLMs have demonstrated impressive capabilities, yet their ability to infer outcomes and causes remains underexplored. To address this, we introduce **NL-Eye**, a benchmark designed to assess VLMs' **visual abductive reasoning skills**. 
NL-Eye adapts the **abductive Natural Language Inference (NLI)** task to the visual domain, requiring models to evaluate the **plausibility of hypothesis images** based on a premise image and explain their decisions. The dataset contains **350 carefully curated triplet examples** (1,050 images) spanning diverse reasoning categories, temporal categories and domains.
NL-Eye represents a crucial step toward developing **VLMs capable of robust multimodal reasoning** for real-world applications, such as accident-prevention bots and generated video verification.

project page: [NL-Eye project page](https://venturamor.github.io/NLEye/)

preprint: [NL-Eye arxiv](https://arxiv.org/abs/2410.02613)

---

## **Dataset Structure**
The dataset contains:
- A **CSV file** with annotations (`test_set.csv`).
- An **images directory** with subdirectories for each sample (`images/`).

### **CSV Fields:**
| Field                          | Type     | Description                                                    |
|--------------------------------|----------|----------------------------------------------------------------|
| `sample_id`                    | `int`    | Unique identifier for each sample.                             |
| `reasoning_category`           | `string` | One of the six reasoning categories (physical, functional, logical, emotional, cultural, or social). |
| `domain`                       | `string` | One of the ten domain categories (e.g., education, technology).     |
| `time_direction`               | `string` | One of three directions (e.g., forward, backward, parallel).                 |
| `time_duration`                | `string` | One of three durations (e.g., short, long, parallel).                  |
| `premise_description`          | `string` | Description of the premise.                               |
| `plausible_hypothesis_description` | `string` | Description of the plausible hypothesis.                        |
| `implausible_hypothesis_description` | `string` | Description of the implausible hypothesis.                      |
| `gold_explanation`             | `string` | The gold explanation for the sample's plausibility.             |
| `additional_valid_human_explanations` | `string` (optional) | Extra human-generated (crowd-workers) explanations for explanation diversity. |

> **Note**: Not all samples contain `additional_valid_human_explanations`.

---

### **Images Directory Structure:**
The `images/` directory contains **subdirectories named after each `sample_id`**. Each subdirectory includes:
- **`premise.png`**: Image showing the premise.
- **`hypothesis1.png`**: Plausible hypothesis.
- **`hypothesis2.png`**: Implausible hypothesis.

## **Usage**
This dataset is **only for test purposes**. 

### Citation
```bibtex
@misc{ventura2024nleye,
  title={NL-Eye: Abductive NLI for Images},
  author={Mor Ventura and Michael Toker and Nitay Calderon and Zorik Gekhman and Yonatan Bitton and Roi Reichart},
  year={2024},
  eprint={2410.02613},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}