Diffusion LLM with Native Variable Generation Lengths: Let [EOS] Lead the Way

TL;DR

This paper introduces a diffusion-based large language model capable of native variable-length text generation by predicting the end-of-sequence token, significantly improving inference speed and flexibility over fixed-length diffusion models.

Contribution

The paper proposes a novel diffusion LLM with native variable generation lengths, enabling more flexible and efficient text generation by predicting the [EOS] token during diffusion.

Findings

30.1x speedup over traditional dLLM inference

2.4x faster than autoregressive models like Qwen and Llama

Achieves higher accuracy and practical inference speed

Abstract

Diffusion-based large language models (dLLMs) have exhibited substantial potential for parallel text generation, which may enable more efficient generation compared to autoregressive models. However, current dLLMs suffer from fixed generation lengths, which indicates the generation lengths of dLLMs have to be determined before decoding as a hyper-parameter, leading to issues in efficiency and flexibility. To solve these problems, in this work, we propose to train a diffusion LLM with native variable generation lengths, abbreviated as dLLM-Var. Concretely, we aim to train a model to accurately predict the [EOS] token in the generated text, which makes a dLLM be able to natively infer in a block diffusion manner, while still maintaining the ability of global bi-directional (full) attention and high parallelism. Experiments on standard benchmarks demonstrate that our method achieves a 30.1x speedup over traditional dLLM inference paradigms and a 2.4x speedup relative to autoregressive models such as Qwen and Llama. Our method achieves higher accuracy and faster inference, elevating dLLMs beyond mere academic novelty and supporting their practical use in real-world applications. Codes and models have been released.