Large Language Models (LLMs), particularly multimodal ones (MLLMs), may unintentionally store sensitive information across text and images, creating security risks; the UnLOK-VQA benchmark and an “attack-and-defense” framework are introduced to evaluate and improve the effectiveness of unlearning techniques in MLLMs.

This worl introduces a new dataset designed for image-text summarization. It is obtained through a self-refinement process involving human annotations and critic model filtering that significantly enhances the summary quality.

Propose methods for debiasing multimodal models involving confounder modeling using information minization and causal debiasing.

Can Sensitive Information be Deleted from LLMs? We show that extraction attacks recover “deleted” knowledge between 18 and 38% of the time! Our attack-and-defense framework includes whitebox+blackbox attacks, and our defense objectives bring down attacks from 38%->2.4%. Our findings suggest that truly deleting sensitive information is a tractable but difficult problem. This result has potentially severe implications for deployment of LLMs in a world where individuals enjoy ownership of their data, a right to privacy, and safety from harmful model outputs.

Propose a novel scene expansion task where we enrich an input seed graph by adding new nodes (objects) and the corresponding relationships.

Propose Overlap BPE (OBPE), an effective modification to the BPE vocabulary generation algorithm which enhances overlap across related languages. Unlike previous studies that dismissed the importance of token-overlap, we show that in the low-resource related language setting, token overlap matters.

Propose a three-stage approach for the task of identifying versions and ordering them correctly in this paper and show that it significantly outperforms state-of-the-art approach adapted for this task, which justifies defining this problem as a novel challenge.

Experiments on multiple real-world benchmark datasets provide validation to our hypothesis that using a related language as pivot, along with transliteration and pseudo translation based data augmentation, can be an effective way to adapt LMs for LRLs, rather than direct training or pivoting through English.