metadata
license: apache-2.0
AIMS Benchmarks
This repository contains the benchmark data and evaluation resources used in AIMS (Adaptive Information Multi-source Steering).
Contents
AIMS_Benchmarks/
├── MME_Benchmark_release_version/ # MME
├── amber/ # AMBER
├── chair_coco/ # CHAIR
├── faithscore/ # FaithScore
└── nltk_3-8-1/ # NLTK resources
The nltk_3-8-1 directory contains the taggers, tokenizers, and corpora required by the evaluation scripts. These resources correspond to NLTK 3.8.1 and are provided for convenient offline evaluation.
Usage
- Download the whole repository:
hf download VisionXLab/AIMS_Benchmarks --local-dir <your_local_path> --repo-type dataset
- Download a specific directory (e.g., AMBER):
hf download VisionXLab/AIMS_Benchmarks --local-dir <your_local_path> --include "amber/**" --repo-type dataset
Acknowledgements
The included benchmarks and resources are collected from their original projects for reproducible evaluation. Please refer to and cite the original works of MME, AMBER, CHAIR, FaithScore, MSCOCO, and NLTK when using the corresponding resources.
Cite Us
@misc{ma2026visualenhancementadaptivemulticontext,
title={Beyond Visual Enhancement: Adaptive Multi-Context Steering to Mitigate LVLM Hallucinations},
author={Shuran Ma and JiaLe Li and Yuxin Dong and Shan Zheng and Qingyun Jiang and Xiang Chen and Qi Zhu and Deyi Ji and Yifan Yang and Jianfeng Pan and Yu Tian and Xue Yang},
year={2026},
eprint={2610.11907},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.11907},
}