X-ITFormer: Balanced Fusion and Multilevel Auditing for Multimodal Time-Series QA
International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)
DOI: 10.14569/ijacsa.2026.0171099 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von Leon Felipe Davis Coropuna und 2 weiteren Autor:innen.
Abstract (Originalsprache)
LLMs help narrow the semantic gap in time-series analysis by translating numerical patterns into natural-language descriptions. However, their adoption in critical domains is hindered by high computational requirements (models of 7B parameters or more) and by the difficulty of verifying whether predictions rely on the physical signal or on spurious textual correlations. To address these limitations, we propose X-ITFormer, a multimodal framework that builds on ITFormer using a frozen 0.5B-parameter Small Language Model (SLM). It introduces Balanced Fusion, a lightweight pre-attentional normalization mechanism using only 1,794 additional trainable parameters, and a post-hoc X-Module for multi-level auditing through (Level 1) attention-based attribution, (Level 2) k-nearest-neighbor case retrieval, and (Level 3) counterfactual generation. On open-ended queries X-ITFormer leads (0.446 vs. 0.345 Q-HEART and 0.399 ECG-LM) but trails on closed-form tasks (Verify: 0.629 vs. 0.913/0.758; Choose: 0.562 vs. 0.626/0.574); these heterogeneous comparisons do not establish equivalence to larger backbones under a shared protocol. The auditing evaluation shows: 1) attention maps stable under Gaussian noise (cosine similarity 0.9998) with above-chance lead alignment (Top-1 22.81% vs. 8.33%); 2) Balanced Fusion reducing the logged counterfactual mean squared error by 26.3% and the optimization iterations by 38.0% (Wilcoxon, p ¡ 0.001) on N = 1000 binary ECG-QA samples, at the cost of increased sensitivity to localized adversarial perturbations; and 3) the retrieval module preserving record-level identity and above-random reference-answer agreement in a question-matched cross-record corpus audit. These audits provide partial evidence of signal dependence, rather than a certification of grounding. On TSQA the model outperforms ChatTime only at L ≥ 256 with the largest advantage at L = 512 (+0.0885); on EngineMT-QA, Balanced Fusion yields directionally consistent but statistically non-significant improvements (p ≥ 0.057, n = 5 seeds) at 0.11% of FiLM parameters. A single forward pass uses 2.19 GB VRAM (B = 1); SFT cost under $0.01 USD and emitted 0.0824 kg CO2eq on ECG-QA. The 0.5B SLM has limited zero-shot reasoning capacity, and performance may degrade on out-of-distribution data.
Zitieren
APA
Coropuna, L. F. D., Quispe, P. M., & Valdivia, A. M. C. (2026). X-ITFormer: Balanced Fusion and Multilevel Auditing for Multimodal Time-Series QA. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.0171099
Vancouver
Coropuna LFD, Quispe PM, Valdivia AMC. X-ITFormer: Balanced Fusion and Multilevel Auditing for Multimodal Time-Series QA. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.0171099
BibTeX
@article{Coropuna2026,
author = {Coropuna, Leon Felipe Davis and Quispe, Percy Maldonado and Valdivia, Ana Maria Cuadros},
title = {X-ITFormer: Balanced Fusion and Multilevel Auditing for Multimodal Time-Series QA},
journal = {International Journal of Advanced Computer Science and Applications},
year = {2026},
volume = {17},
number = {10},
doi = {10.14569/ijacsa.2026.0171099}
}
EUVIMED