AI models can stop drawing something while still recognizing it

Xue and colleagues find that suppressing image generation does not reliably erase visual recognition, and report better joint forgetting with an approach targeting both capabilities.

Chinese Tech

Zhongkai Xue · Junkai Chen · Mingcong Lei · Rumo Zhang · Yixiao Liu · Zhimin Sun · +4 more

Institute of Automation, Chinese Academy of Sciences · University of Chinese Academy of Sciences · ByteDance · Shanghai Jiao Tong University · The Chinese University of Hong Kong, Shenzhen

Research Digest·4 November 2026·2 min read

Illustration: AI models can stop drawing something while still recognizing it

In one line

UniForget helps unified multimodal models forget target concepts in both image generation and visual recognition better than single-pathway methods.

Before you read this

Three ideas this paper builds on

machine unlearningRemoving specific knowledge, behaviors, or concepts from a trained model while preserving overall capability.13 papers →unified multimodal understanding and generationA single end-to-end model processes and generates multiple data types (e.g., images, text) with equal competence.26 papers →benchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.302 papers →

Read it as

A model that can both make pictures and recognize them may still know a concept after an edit removes it from only one task. Concept unlearning means changing a trained model so it no longer produces or recognizes chosen content, while retaining other abilities.

The team introduces UUBench, a that tests both image generation and visual question answering, or answering questions about pictures. They also develop UniForget, which combines model edits for generation and understanding.

UniForget reaches joint forgetting rates of 80.0% and 92.7% on two tested models, outperforming the evaluated alternatives. The results also show a trade-off: removing a concept can harm other capabilities.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

machine unlearning unified multimodal understanding and generation benchmark low-rank adaptation