Reproduce “Breaking the Factorization Barrier in Diffusion Language 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
Ian Li, Zilei Shao, Benjie Wang, Rose Yu, Guy Van den Broeck, Anji Liu
Coupled Discrete Diffusion (CoDD) enhances discrete diffusion language models by integrating a lightweight, tractable probabilistic circuit to model complex joint dependencies among simultaneously predicted tokens. This approach achieved up to +22.44% accuracy gains on GSM8K in few-step generation scenarios while adding only 4-5% additional inference latency.
Key insights
- The
- factorization barrier
- is a structural misspecification in the dLLM's output distribution, not an inherent limitation of the neural network's capacity to learn dependencies.
- Multiplicative composition of a simple factorized potential with an expressive structural prior allows for complex joint dependency modeling without prohibitive parameter explosion.
- Probabilistic Circuits offer a computationally efficient and tractable means to capture intricate joint dependencies and perform inference, including exact partition function computation, for discrete diffusion models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.