New AI Research Proposes R²-dLLM Framework to Reduce Redundancy in Diffusion Large Language Models

A research paper titled "R²-dLLM: Accelerating Diffusion Large Language Models via Spatio-Temporal Redundancy Reduction" has been released on arXiv, identified as arXiv:2604.18995v1. This study tackles the significant inference latency that hinders the effective implementation of Diffusion Large Language Models (dLLMs), which present a viable alternative to autoregressive generation by facilitating parallel token predictions. The authors identified that much of this inefficiency arises from repetitive redundancy in the decoding phase, including spatial redundancy from confidence clusters and positional ambiguity, along with temporal redundancy from remasking already stabilized predictions. To address these issues, they introduce R²-dLLM, a comprehensive framework aimed at minimizing decoding redundancy during both inference and training. For inference, it offers training-free decoding rules to consolidate local confidence and token predictions, finalizing stable tokens to eliminate unnecessary decoding steps. Additionally, they suggest a redundancy-aware supervised fine-tuning method to bolster the model's redundancy reduction capabilities during training. This paper falls under the cross-announcement category on arXiv.