Predictive advantage of temporal alignment

Establish whether temporal alignment of coherence and diversity trajectories provides a predictive advantage over endpoint diversity, trajectory means, or prompt-relative summary scores for predicting human quality ratings of machine-generated continuations.

Background

The study compares dynamic time-warping scores with simpler endpoint and summary-based measures. Although diversity-based temporal alignment correlates positively with human ratings, the paired comparisons have intervals that include zero and therefore do not establish that temporal alignment is superior.

The unresolved issue is whether modeling the temporal shape of coherence and diversity contributes predictive information beyond aggregate or endpoint measurements.

References

For RQ1, these comparisons leave the predictive advantage of temporal alignment unresolved.

— Reference-Based Analysis of Coherence and Diversity in Open-Ended Text Generation  (2609.28080 - Arias, 23 Sep 2026) in Section 5, subsection “RQ1: Temporal and Summary Associations”

These comparisons do not identify a preferred distribution or domain on held-out data.

— Reference-Based Analysis of Coherence and Diversity in Open-Ended Text Generation  (2609.28080 - Arias, 23 Sep 2026) in Section 5, subsection “RQ3: Reference-Distribution Likelihood”