DARE: Diffusion Language Model Activate Reuse for Efficient Inference

Diffusion Large Language Models (dLLMs) have recently emerged as a compelling alternative to auto-regressive (AR) language models, offering a larger expressive field and significant potential for parallel generation and inference speedup. However, existing open-source dLLMs remain in an early stage of development, lagging behind their AR counterparts in both efficiency and task quality. In this work, we identify and exploit an underexplored property of dLLMs: the high degree of token-wise redundancy in their bi-directional self-attention layers. We observe that self-attention activations are strongly correlated across tokens, and that the temporal change in query representations can be used to predict redundancy in the corresponding key, value, and output activations. Building on this insight, we introduce DARE (dLLM Activation Reuse for Efficient Inference) and its two complementary reuse mechanisms: K-DARE, which selectively reuses cached key-value (KV) activations and O-DARE, which reuses output activations to reduce redundant computation while preserving model quality. Empirically, DARE achieves up to a 1.20× reduction in per-layer latency and reuses as much as 87% of blockwise attention activations, all with negligible quality degradation across reasoning and code-generation benchmarks. Both reuse mechanisms, K-DARE and O-DARE, maintain highly competitive generation quality with average performance drops of only 2.0% and 1.2%, respectively. When combined with existing acceleration techniques such as prefix caching and Fast-dLLM, our method delivers additive performance gains without requiring any model retraining. These results highlight token-wise reuse as a principled and effective strategy for improving the computational efficiency of diffusion-based LLMs, enabling scalable inference while preserving generation fidelity.