FourierSampler: Unlocking Non-Autoregressive Potential in Diffusion Language Models via Frequency-Guided Generation
This work addresses a bottleneck in non-autoregressive language generation for AI researchers, offering a novel decoding strategy to enhance model efficiency and output quality.
The paper tackles the problem of positional bias in diffusion language models by analyzing their spectral characteristics and proposing FourierSampler, a frequency-guided generation method that improves performance by 20.4% on LLaDA1.5-8B and 16.0% on LLaDA-8B-Instruct, surpassing autoregressive models like Llama3.1-8B-Instruct.
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.