Hierarchical Manual Parsing and Order–Spectral Metrics for Structural Evaluation of Multimodal Instructions
Year: 2026
Type: conference-paper
Abstract: Vision-Language Models (VLMs) are increasingly applied to multimodal assistance in consumer electronics, yet procedural manuals require structural coherence beyond semantic similarity. This work proposes a unified framework that combines hierarchical manual extraction with trajectory and spectral-based structural evaluation. Instructions and responses are modeled as ordered embedding sequences, enabling quantification of ordering fidelity, multiscale distortion, and multimodal consistency. Experiments on 36 sections of the Alienware 17 R5 Service Manual under controlled perturbations show that rank-based and wavelet metrics detect procedural distortions largely invisible to conventional measures such as BLEU and BERTScore. These results demonstrate that structural signal modeling provides a principled and model-agnostic approach for evaluating multimodal instructional quality in embedded LLM applications.
Authors: Diego Amoedo, Ruan Belém, Edma Mattos, Elton Alencar, Victória Guimarães, Pedro Matias, Emanuel Oliveira, Luca Naja, Matheus Castro, Pedro Oliveira, Matheus Uchoa, Cristian Maia, Lucas Botinelly, Rosiane Brito, Bruno Cardoso, Laura Leite, Lucas Pessoa, Alexandre Miranda, Agemilson Pimentel, Rômulo Fabricio, Celso B. Carvalho, Waldir S. S. Júnior
Institutions: Universidade Federal do Amazonas, TPV Technology (China)
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204127705
