PPU-Bench: Real-World Multimodal Benchmark for Personalized Partial Unlearning

Jiahui Guang, Zexun Zhan, Zhenlin Xu, Cuiyun Gao, Haiyan Wang, Jing Li, Zhaoquan Gu, Yanchun Zhang · 2026

Read PPU-Bench: Real-World Multimodal Benchmark for Personalized Partial Unlearning by Jiahui Guang, Zexun Zhan, Zhenlin Xu, Cuiyun Gao, Haiyan Wang, Jing Li, Zhaoquan Gu, Yanchun Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-08800.txt.

Description

Jiahui Guang and seven coauthors introduce PPU-Bench, a benchmark for deleting personal knowledge from vision language models in the fine grained way real requests demand. Earlier unlearning tests used synthetic injections or wiped whole subjects, missing cases where one fact about a person must go while neighbors stay. Their set holds twenty four thousand samples built from existing knowledge of five hundred public figures across three graded settings.

Experiments show complete deletion often suppresses visual identity instead of facts, while selective settings expose sharp forget versus retain trade offs and attacks that cross images and prompts. Their boundary aware optimization improves two methods' handling of those inner boundaries. Privacy tooling gains a realistic yardstick.

Catalog record

CREATORJiahui Guang, Zexun Zhan, Zhenlin Xu, Cuiyun Gao, Haiyan Wang, Jing Li, Zhaoquan Gu, Yanchun Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE75 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.08800

Browse the full library · About the Tower of Babel