Improved Token Acceptance Estimators

Background: Developing faster decoding methods for large language models is crucial for practical deployment, especially for computationally intensive models like block-diffusion LMs.

Question / Future Work: Investigate alternative, potentially more effective, token acceptance estimators for predicting the expected accepted prefix length ($\hat{K}$) beyond margin-based and entropy-based methods, as better estimation accuracy might lead to further downstream decoding performance gains.