Reproduce “Flexible-length Text Infilling for Discrete Diffusion Models” 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
Andrew Zhang, Anushka Sivakumar, Chiawei Tang, Chris Thomas
Virginia Tech researchers introduce DDOT (Discrete Diffusion with Optimal Transport Position Coupling), a method that allows discrete diffusion models to perform flexible-length text infilling by jointly denoising token values and continuous positions. This system outperforms naive diffusion baselines and achieves performance on par with state-of-the-art non-autoregressive models, maintaining a nearly 100% success rate on block infilling tasks due to its novel Optimal Transport coupling.
Key insights
- Flexible text infilling requires jointly learning both what tokens to generate and where to place them in a fluid, continuous manner.
- A simple, order-preserving Optimal Transport coupling is crucial for making continuous position diffusion tractable by reducing the combinatorial complexity of permutations and preserving the learning signal.
- Maintaining the relative order of prompt and response tokens during diffusion is essential for semantic coherence and preventing invalid text generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.