Title: Fast-dLLM++: Leveraging Fréchet Profile Decoding to Accelerate Diffusion LLM Inference

Abstract:

While diffusion large language models offer the potential for parallel token generation, their inference speeds are currently limited by the challenge of determining which masked tokens can be securely committed simultaneously. The existing Fast-dLLM approach mitigates this issue through confidence-guided parallel decoding combined with KV caching. However, its underlying decoding theory relies on a homogeneous high-confidence assumption, which effectively constrains each candidate set to the confidence level of its weakest selected token. We contend that this limitation results in suboptimal performance, as actual decoding steps display heterogeneous confidence profiles.

To address this, we introduce Fast-dLLM++, a training-free enhancement that implements Fréchet profile decoding. Unlike previous methods that rely on a single worst-case confidence value, this approach selects parallel commit sets based on the entire sorted confidence profile. This new rule serves as a heterogeneous-confidence generalization of Fast-dLLM’s factor selector; it exactly replicates the original rule when confidences are equal but introduces a provable heterogeneity bonus when selected tokens possess uneven confidence levels.

Fast-dLLM++ requires no modifications to the model architecture, diffusion process, or cache implementation, allowing it to function as a seamless drop-in replacement for existing Fast-dLLM decoding systems. Empirical evaluations conducted on the LLaDA-8B model across datasets including GSM8K, MATH, HumanEval, and MBPP demonstrate that these theoretical advantages yield tangible performance improvements. By identifying safe parallelism opportunities that weakest-token rules overlook, profile-aware selection enhances the accuracy-throughput trade-off, delivering up to 37% higher throughput with comparable accuracy. An anonymous code release is available at https://github.com/Ringo-Star/FastdLLM_plusplus.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC