Reference-Based Framework for Evaluating Text Coherence and Diversity

A new evaluation framework measures text generation quality by comparing model outputs against human reference distributions and trajectories. Experiments show that diversity-based alignment and mean-based comparisons correlate with human quality ratings, offering structured alternatives to standard metrics.