Knowledge Editing in Masked Diffusion Language Models

Abstract: Knowledge editing aims to update or correct factual knowledge in a LLM. A widely used approach, locate-then-edit, does this in two steps: it first localizes a fact within the model, then edits the weights there. To date, such methods have been developed exclusively on autoregressive models (ARMs). Whether their underlying assumptions hold for masked diffusion models (MDMs), which model text bidirectionally and generate by iterative denoising rather than next-token prediction, remains an open question. We address it by transferring locate-then-edit to MDMs and comparing two MDMs (LLaDA, Dream) with two ARMs (LLaMA, Qwen) at matched scale. Our central finding has two parts. First, where an edit is applied transfers across paradigms: causal tracing highlights the same early-to-mid-layer MLP at the last subject token in both, and editing is most effective there. Second, this shared location does not guarantee a shared outcome. Single-token edits succeed in both, but as targets grow longer, editing degrades systematically in the MDMs but not the ARMs. The failure stems from how the edited fact is generated: producing a multi-token target requires passing through partially unmasked intermediate states for which the edit was never optimized. Guided by this diagnosis, we introduce a simple correction that optimizes the edit for these states, substantially restoring multi-token performance.