---
format: "aidr-story-markdown/v1"
id: "1d099c8ea4eaaa76c3e2a61880de582e6e8d889d77c5fa6834da5b3d092728ea"
canonical_url: "https://aidr.today/1d099c8e?lang=en"
title: "Rethinking Supervision Targets for Multimodal LLM Unlearning: Less Private Information, Reduced Distribution Shift"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-10T05:00:00.000Z"
category: "Research"
topics: ["safety"]
source_urls: ["https://arxiv.org/abs/2611.09939"]
summary: "Multimodal Large Language Models can memorize and reproduce sensitive information from their training data. Supervision targets guiding model unlearning often trade forgetting effectiveness for model utility. FFN-Perturbed Self-Distillation is proposed, starting from the teacher's original output distribution as a zero-shift reference."
---

# Rethinking Supervision Targets for Multimodal LLM Unlearning: Less Private Information, Reduced Distribution Shift

> [Open the canonical story](<https://aidr.today/1d099c8e?lang=en>)

**Published:** 2026-11-10T05:00:00.000Z
**Category:** Research
**Topics:** safety

## Summary

Multimodal Large Language Models can memorize and reproduce sensitive information from their training data\. Supervision targets guiding model unlearning often trade forgetting effectiveness for model utility\. FFN\-Perturbed Self\-Distillation is proposed, starting from the teacher's original output distribution as a zero\-shift reference\.

## Sources

- [Story source](<https://arxiv.org/abs/2611.09939>)