Towards Benign Memory Forgetting for Selective Multimodal Large Language Model Unlearning

A team of researchers has proposed a new framework called the Sculpted Memory Forgetting Adapter (SMFA) to enable large language models to selectively forget sensitive information without crippling their general capabilities, according to a paper posted on arXiv [1]. Multimodal large language models (MLLMs) can inadvertently memorize privacy-sensitive information during training [1]. While existing unlearning methods can remove such content, they often severely degrade the model's foundational capabilities, such as general image understanding [1]. This shortfall motivated the investigation into what the authors term "benign memory forgetting" — the precise removal of targeted, privacy-sensitive knowledge while rigorously preserving unrelated capabilities [1]. The paper was submitted to arXiv, the open-access repository of electronic preprints that is not peer-reviewed, in November 2025 and revised in June 2026 [1][6]. To measure progress toward this goal, the researchers introduced S-MLLMUn Bench, described as the first benchmark designed to jointly and quantitatively assess an unlearning method's efficacy in knowledge erasure and the preservation of image understanding [1]. The proposed SMFA framework confines forgetting to designated memory regions, maintaining overall model performance [1]. It works by initially fine-tuning the model to replace sensitive outputs with refusals, generating a memory forgetting adapter, and then applying a retaining anchor-guided masking mechanism that safeguards unrelated knowledge [1]. Extensive experiments on S-MLLMUn Bench demonstrated that existing methods fail to achieve benign forgetting, whereas SMFA successfully achieved targeted knowledge erasure without compromising the model's foundational visual capabilities [1]. The code and data for the project have been made available on GitHub [1]. The paper was authored by Zhen Zeng and colleagues [1]. arXiv, which began on August 14, 1991, passed the two-million-article milestone by the end of 2021 and now receives about 24,000 submissions per month [6]. The platform hosts papers across mathematics, physics, computer science, and other fields, and has become a standard venue for disseminating early research findings before formal peer review [6]. The SMFA paper appears in the Artificial Intelligence section of the computer science category [1].