Researchers have published multiple approaches to address the high inference latency that limits deployment of diffusion large language models (dLLMs), which generate text by iteratively denoising masked segments rather than using autoregressive generation.
According to arxiv.org, the $R^{2}$-dLLM framework reduces decoding redundancy from both inference and training perspectives. The research identifies spatial redundancy from confidence clusters and positional ambiguity, plus temporal redundancy from repeatedly remasking stabilized predictions. The method “consistently reduces the number of decoding steps by up to 88% compared to existing decoding strategies, while maintaining competitive generation quality,” according to the paper.
In a separate approach, Fast-dLLM++ introduces “Fréchet profile decoding” that selects parallel commit sets from full sorted confidence profiles rather than single worst-case confidence values, according to arxiv.org. The paper states this achieves “up to 37% higher throughput at comparable accuracy” on benchmarks including GSM8K and MATH with the LLaDA-8B model.
A third technique, dLLM-Cache, addresses the incompatibility of traditional Key-Value caching with dLLMs’ bidirectional attention. According to arxiv.org, the method combines “long-interval prompt caching with partial response updates guided by feature similarity,” achieving “up to 9.1x FLOPs reduction on LongBench-HotpotQA while maintaining competitive output quality.” The paper was accepted by ICML 2026.
All three methods are described as training-free approaches that maintain model performance while significantly reducing computational costs.