MEA-Benchmark

Benchmark for MEA: A Reward-Driven Multi-Agent System for Faithful Model Explanations.

• 📄 Paper: arXiv:2610.02480
• 💻 Code: github.com/AikyamLab/xai-agent

MEA-Benchmark evaluates explanations of neural network models across three modalities (tabular, text, vision) with ten question types (Q1–Q10), spanning feature attribution, counterfactual reasoning, and spurious feature detection. Each question type is paired with a perturbation-based faithfulness metric (see the paper and code).

Dataset Structure

The dataset is available in two forms:

1. viewer/{train,test}.jsonl: the default config, one unified-schema table per split (17,344 rows in total: 13,868 train / 3,476 test). This is what the Dataset Viewer and load_dataset use.
2. {split}/{modality}/{dataset}_{model}_{q_type}.json: the original per-file JSON, which is what the MEA code reads.

from datasets import load_dataset
ds = load_dataset("EstherrrCheng/mea-benchmark")  # splits: train, test

{split}/
├── tabular/   adult_{2layernn,tabnn}_q{1..10}.json, cancer_{2layernn,tabnn}_q{1..10}.json
├── text/      imdb_{2layernn,cnn}_q{1..10}.json,    snli_{2layernn,cnn}_q{1..10}.json
└── vision/    cub_{resnet,densenet}_q{1..10}.json,  stl10_{resnet,densenet}_q{1..10}.json

Fields (default config)

All columns are strings except q_type. Fields whose structure differs between question types (dicts, lists) are stored as JSON strings; use json.loads to decode them. A field that does not apply to a row is an empty string.

Field	Description
source_file	Original JSON file this row came from
modality	tabular, text, or vision
dataset	adult, cancer, imdb, snli, cub, or stl10
model	Model architecture (2layernn, tabnn, cnn, resnet, densenet); comma-separated if several
q_type	Question type, integer 1–10
q	The XAI question prompt
example	Raw input example
row_no	Sample index (JSON; a list for multi-instance questions)
features	Input features or tokens (JSON)
target	Ground-truth label (JSON)
predicted	Model prediction (JSON)
image_path, image_id	Input image (vision, single-instance)
image_paths, image_indices	Input images (vision, multi-instance questions; JSON)
instance_A, instance_B, pair_id	Instance pair (Q4 and pairwise questions; JSON)

Question Types

Type	Question
Q1	Which part of the input was most responsible for the model's prediction?
Q2	Which part of the input was least responsible for the model's prediction?
Q3	Which parts distinguish the prediction from the next-best alternative?
Q4	Why are instances A and B given different predictions? (contrastive instances)
Q5	If a certain part is masked, would the prediction change?
Q6	How should the instance change to flip the prediction to the expected class?
Q7	If one important part is removed or changed, how would the prediction change?
Q8	Is there an irrelevant (spurious) part causing the model's wrong prediction?
Q9	What shared feature makes multiple misclassified inputs difficult?
Q10	Why are two similar instances given different predictions (one correct, one wrong)?

Citation

@article{cheng2026mea,
  title         = {{MEA}: A Reward-Driven Multi-Agent System for Faithful Model Explanations},
  author        = {Cheng, Yuyang and Ravi, Raghav Kaushik and Sridhar, Srivarshinee and Saha, Sriparna and Ghosh, Akash and Agarwal, Chirag},
  journal       = {arXiv preprint arXiv:2610.02480},
  year          = {2026},
  eprint        = {2610.02480},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}