FourierSampler: Unlocking Non-Autoregressive Potential in Diffusion Language Models via Frequency-Guided Generation

TL;DR

FourierSampler introduces a frequency-guided decoding method for diffusion language models, leveraging spectral analysis to improve global structure and local detail generation, outperforming existing strategies and autoregressive models.

Contribution

This work is the first to analyze diffusion language models in the frequency domain and proposes FourierSampler, a novel frequency-guided decoding technique that enhances generation quality.

Findings

Achieves 20.4% relative improvement on LLaDA1.5-8B

Achieves 16.0% relative improvement on LLaDA-8B-Instruct

Surpasses autoregressive models like Llama3.1-8B-Instruct

Abstract

Despite the non-autoregressive potential of diffusion language models (dLLMs), existing decoding strategies demonstrate positional bias, failing to fully unlock the potential of arbitrary generation. In this work, we delve into the inherent spectral characteristics of dLLMs and present the first frequency-domain analysis showing that low-frequency components in hidden states primarily encode global structural information and long-range dependencies, while high-frequency components are responsible for characterizing local details. Based on this observation, we propose FourierSampler, which leverages a frequency-domain sliding window mechanism to dynamically guide the model to achieve a "structure-to-detail" generation. FourierSampler outperforms other inference enhancement strategies on LLADA and SDAR, achieving relative improvements of 20.4% on LLaDA1.5-8B and 16.0% on LLaDA-8B-Instruct. It notably surpasses similarly sized autoregressive models like Llama3.1-8B-Instruct.