DARE: Boosting Diffusion LLM Inference via Activation Reuse

Researchers have introduced DARE (Diffusion Language Model Activation Reuse), a method to accelerate inference in diffusion large language models (dLLMs) by exploiting token-wise redundancy in self-attention. The approach comprises two mechanisms: DARE-KV reuses cached key-value activations, while DARE-O reuses output activations, reducing redundant computation without significant quality loss. Experiments show up to 1.20x per-layer latency reduction and reuse of up to 87% of attention activations. The work addresses the current immaturity of open-source dLLMs compared to auto-regressive models, offering potential for faster parallel generation. The paper is available on arXiv under identifier 2605.08134.