AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

SAUCE: Selective Concept Unlearning in Vision-Language Models with Sparse Autoencoders

Jiahui Geng, Qing Li

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 6 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use SAUCE SAE intervention instead of PO (second-best baseline)LLaVA-v1.5-7B · Concrete concept unlearningUAg: 74.6% -> 86.3%Improved
use SAUCE instead of PO on abstract conceptsLLaVA-v1.5-7B · Abstract concept unlearningUAd: 64.6% -> 87.4%Improved
use SAUCE instead of PO for in-domain retentionLLaVA-v1.5-7B · Concrete concept unlearningIRA: 84.7% -> 90.6%Improved
use SAUCE instead of PO on abstract tasks for cross-domain retentionLLaVA-v1.5-7B · Abstract concept unlearningCRA: 84.8% -> 83.4%Hurt
use SAUCE instead of PO for model utilityLlama-3.2-11B-Vision-Instruct · Abstract concept unlearningMME: 93.1% -> 90.7%Hurt
switch the VLM from LLaVA-v1.5-7B to Llama-3.2-11B-Vision-InstructSAUCE · Concrete concept unlearningUAg: 86.3% -> 93.8%Improved
move from concrete to abstract conceptsSAUCE · LLaVA-v1.5-7BUAg: 86.3% -> 83.1%Hurt
move from concrete to abstract conceptsKL · LLaVA-v1.5-7BUAg: 70.5% -> 59.4%Hurt
transfer SAE trained on LLaVA-v1.5-7B to LLaVA-v1.5-13B (same vision encoder)SAUCE · Concrete concept unlearningUAg: 86.3% -> 84.3%No clear effect
transfer SAE trained on LLaVA-v1.5-7B to LLaVA-v1.6-Vicuna-7B (different vision encoder)SAUCE · Concrete concept unlearningMME: 90.1% -> 40.7%Hurt
increase number of concepts unlearned simultaneouslySAUCE · LLaVA-v1.5-7B multi-conceptutility scoreMixed
apply Chinese translation and 4-bit quantization attacksSAUCE · LLaVA-v1.5-7B concreteUAd: decline of 5%Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy