Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference

In three linesFast-dLLM++ improves diffusion LLM inference by replacing homogeneous confidence token selection with Fréchet profile decoding. Training-free, it exploits heterogeneous confidence profiles to parallelize more tokens safely, achieving up to 37% higher throughput on GSM8K, MATH, HumanEval, and MBPP with LLaDA-8B while maintaining accuracy.

Summary generated by Claude — human-verified