<ul><li>Diffusion LLMs are a promising alternative to autoregressive LLMs, providing enhanced runtime efficiency.</li><li>Existing diffusion models lack the ability to enforce user-specified formal constraints, hindering their reliability for tasks requiring structured outputs.</li><li>DINGO is proposed as a dynamic programming-based constrained decoding strategy to address this limitation in diffusion LLMs.</li><li>DINGO allows sampling of high-probability output strings while adhering to user-specified regular expressions, showing significant improvements in benchmarks.</li></ul>

DINGO: Constrained Inference for Diffusion LLMs

Discover more