Reproduce “Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yeongbin Seo, Dongha Lee, Jaehyung Kim, Jinyoung Yeo

Researchers at Yonsei University developed Convolutional decoding (Conv) and Rejecting Rule-based Fine-Tuning (R2FT) to address limitations in diffusion language models, specifically the "long decoding-window problem" and "time-interval expansion problem." Their approach achieves state-of-the-art performance among diffusion LMs in open-ended text generation and generates text approximately 3x faster than autoregressive models at reduced decoding steps.

Key insights

  • The "long decoding-window problem" is a critical bottleneck for fluency and coherence in diffusion LMs, causing unaligned token predictions far from the input context.
  • Current semi-autoregressive methods for diffusion LMs compromise both the potential for high speed and bidirectional generation, failing to fully leverage the core advantages of diffusion architectures.
  • Preference optimization can be effectively achieved using rule-based negative samples, offering a cost-efficient alternative to human feedback for mitigating specific undesirable generation patterns like repetition.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.