PEST: Parameter Efficient Steering of Blackbox VLMs via Agentic Few-shot Alignment for Hateful Meme Moderation

🎉 ACM Multimedia (ACM MM) Main Track Paper Accepted!

Figure 1. Overview of the PEST framework.

## 🔎 TL;DR

PEST introduces a novel framework for generalizable hateful meme moderation that leverages lightweight task-specific multimodal agents to steer powerful black-box vision-language models through agentic few-shot alignment.

Going beyond conventional hate detection, PEST jointly performs classification, explanation, and intervention generation, bringing these complementary tasks together within a unified content moderation pipeline.

Key Contributions

• Introduction of the first unified framework for simultaneous hateful meme classification, explanation, and intervention in a low-resource setup.

• Development of PEST: lightweight multimodal steering agents that effectively guide much larger black-box VLMs without modifying their parameters, enabling efficient and scalable deployment.

• Curation of a new benchmark dataset with high-quality explanation and intervention annotations for end-to-end evaluation of multimodal content moderation.

• Establishment of state-of-the-art performance across multiple benchmarks while maintaining a practical, low-resource framework suitable for real-world deployment.

⚠️ Warning: Contains potentially hateful and offensive content.

📋 Fields

Column	Description
img	Embedded meme image
img_path	Relative path to the meme image
label	Binary classification label
explanation	Natural-language explanation
intervention	Generated moderation intervention

🤗 Loading the Dataset

The dataset can be loaded directly from the Hugging Face Hub using the 🤗 Datasets library.

⚙️ Installation

pip install datasets pillow

📦 Load the Dataset

from datasets import load_dataset

dataset = load_dataset("swainsubhankar/PEST")

print(dataset)

The dataset consists of three predefined splits:

DatasetDict({
    FHM_test: Dataset({
        features: ['img', 'img_path', 'label',
                   'explanation', 'intervention']
    })
    FHM_train_explanation: Dataset({
        features: ['img', 'img_path', 'label',
                   'explanation', 'intervention']
    })
    MAMI_test: Dataset({
        features: ['img', 'img_path', 'label',
                   'explanation', 'intervention']
    })
})

📑 Access Individual Splits

fhm_test = dataset["FHM_test"]

fhm_train = dataset["FHM_train_explanation"]

mami_test = dataset["MAMI_test"]

🖼️ Display an Image

example = fhm_train[0]

print("Label:", example["label"])
print("Explanation:", example["explanation"])
print("Intervention:", example["intervention"])

example["img"].show()

📚 Citation

If you use this dataset or framework in your research, please cite:

@misc{rizwan2026seeexplainintervenefewshot,
      title={See, Explain, and Intervene: A Few-Shot Multimodal Agent Framework for Hateful Meme Moderation},
      author={Naquee Rizwan and Subhankar Swain and Paramananda Bhaskar and Gagan Aryan and Shehryaar Shah Khan and Animesh Mukherjee},
      year={2026},
      eprint={2601.04692},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.04692},
}