Video Quality Assessment Principles Applied to Text: Spectral Metrics for LLM Evaluation

View at publisher Summarise with VIRAT
Authors: Diego A. Amoedo, Edma V.C.U. Mattos, Elton Alencar, Victória Guimarães, Pedro Victor Matias, Emanuel Oliveira, Luca Naja, Matheus Castro, Pedro Roberto OLIVEIRA, Matheus S. Uchôa, Cristian Maia, Lucas Botinelly
+10 more Rosiane Brito, Bruno Cardoso, Laura Leite, Lucas Pessoa, Alexandre Miranda, Agemilson Pimentel, Ruan J.S. Belem, Rômulo Fabrício, Celso B. Carvalho, Waldir S. S. Júnior
Cited by: 0
References: 15

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Large Language Models (LLMs) have shown remarkable advances in text generation, but assessing the quality of their responses remains a challenge. Traditional metrics (BLEU, ROUGE, BERTScore) fail to capture discourse-level coherence and fluency. Inspired by Video Quality Assessment (VQA), this paper proposes a framework for evaluating AI-generated text by modeling sequential discourse as temporal signal patterns in the embedding space. The method analyzes both local distortions and global continuity through frequency-domain descriptors such as spectral smoothness and divergence. Experiments with human-written, industrial, and synthetic datasets revealed a consistent correlation between semantic coherence and spectral regularity, achieving up to 93% accuracy in distinguishing human from model-generated text. The results indicate that VQA-inspired spectral metrics can provide a perceptually aligned basis for automated evaluation of language models in consumer-electronics applications.

Fields of Study & Subjects

Topic Modeling Image and Video Quality Assessment Natural Language Processing Techniques

Publication details

Journal / source
International Conference on Consumer Electronics
Pages
1–3
Publication date
3 February 2026
Language
EN