<ul><li>Study focuses on relaxing the problem of coupling probability distributions by generating samples and declaring accept if any sample matches another distribution's sample.</li><li>Proposed a novel method for generating samples, building upon Gumbel-max sampling approach, while establishing a lower bound on the acceptance probability known as the list matching lemma.</li><li>Developed a new mechanism for multi-draft speculative sampling that competes well with existing baselines in various language tasks, ensuring a degree of drafter invariance and providing a theoretical lower bound on token level acceptance probability.</li><li>Introduced a distributed lossy compression technique utilizing the generalized Gumbel-max sampling, demonstrating significant improvements in experiments involving synthetic Gaussian sources and the MNIST image dataset.</li></ul>

List-Level Distribution Coupling with Applications to Speculative Decoding and Lossy Compression

Discover more