VNA_Benchmark

Data can be downloaded here: https://figshare.com/articles/dataset/Multimodal_LLMs_Struggle_with_Basic_Visual_Network_Analysis_a_Visual_Network_Analysis_Benchmark/25938448

Visual Network Analysis Benchmark

This repository contains the data and labels for “Multimodal LLMs Struggle with Basic Visual Network Analysis: a VNA Benchmark”

All graphs were generated using the networkx and netgraph python libraries

Data

The repository contains 4 folders:

• degree
• structural_balance
• connected_components
• labels

the first three folders contain the graph images generated to evaluate GPT-4 and LLAVA and the final folder contains ground-truth labels for each of these graphs. Details on each evaluation are provided below

Degree

Degree labels can be found in ‘labels/degree_labels.json’. This file is most easily read with pd.read_json(). It contains the following fields:

• num_nodes: total number of nodes in the graph
• max_degree: maximum degree in the graph. Identical for letter and number graphs
• nodes_with_max_degree: a list of nodes with maximal degree.
• letter_nodes_with_max_degree: same as the previous column, but with integers mapped to letter IDs.
• file: name of the file in either degree/letter_nodeIDs or degree/number_nodeIDs the graph corresponds to.

Structural Balance

Labels are present in both image and the folder names. All triads in ‘structural_balance/balanced_triads’ are balanced and all triads in ‘structural_balance/imbalanced_triads’ are imbalanced.

Connected Components

‘connected_component_labels.csv’ contains 3 columns:

• component_count: number of components
• isolate_count: number of isolates
• file: name of the file the labels correspond to

Citation

@article{williams2024multimodal,
  title={Multimodal LLMs Struggle with Basic Visual Network Analysis: a VNA Benchmark},
  author={Williams, Evan M and Carley, Kathleen M},
  journal={arXiv preprint arXiv:2405.06634},
  year={2024}
}