TimeROME-DLM: Temporal Causal Tracing and Low-Rank Inference-Time Knowledge Editing for Masked Diffusion Language Models

TimeROME-DLM is a novel inference-time knowledge-editing framework for masked diffusion language models (MDLMs) that operates without training or gradient updates. It employs a Temporal Indirect Effect (TIE) causal-tracing protocol to effectively identify influential coordinates for knowledge editing, coupled with a low-rank residual edit memory that applies updates during diffusion. This approach significantly reduces forget-set log-probability by approximately 83 nats on the TOFU forget01 benchmark using LLaDA-8B-Base while maintaining utility in retain-set log-probability, achieving a four- to fourteen-fold speedup in inference with no additional VRAM, thereby enhancing efficiency for practitioners working with MDLMs.