text-evaluation

This paper introduces DiffScore, a text evaluation framework based on Masked Large Diffusion Language Models that addresses positional bias in autoregressive scoring by using masked reconstruction.