Follow the Latent Roadmap: Navigating Revocable Decoding for Diffusion LLMs with Anchor Tokens

The article presents ASRD (Anchor Supervised Revocable Decoding), a novel framework for improving decoding quality in Diffusion Large Language Models (dLLMs) by using Anchor Tokens to manage uncertainty in token generation. This approach addresses issues of Error Propagation and Local Error Reinforcement by distinguishing between trusted and uncertain tokens, utilizing mechanisms such as Anchor-Guided Generation and Anchor-Perturbed Verification. Experimental results show that ASRD enhances accuracy by up to 6.4% and increases inference throughput by up to 7.2 times, making it a significant advancement for practitioners working with LLMs in scenarios requiring high-quality output and efficiency.