Machine Unlearning for Masked Diffusion Language Models
Researchers have introduced Masked Diffusion Unlearning (MDU), a novel framework designed to remove specific knowledge from Masked Diffusion Language Models (MDLMs). Unlike traditional autoregressive models, MDLMs generate text in parallel by denoising masked positions. MDU adapts the unlearning process to this diffusion-based generation, aiming to shift model predictions away from specific learned information while maintaining utility. Experiments demonstrate MDU's effectiveness in unlearning for MDLMs, outperforming existing methods. AI
IMPACT Introduces a new technique for controlling knowledge within diffusion-based language models, potentially improving privacy and safety.