๐Ÿ“ž +91-7667918914 | โœ‰๏ธ ijarcce@gmail.com
International Journal of Advanced Research in Computer and Communication Engineering
International Journal of Advanced Research in Computer and Communication Engineering A monthly Peer-reviewed & Refereed journal
ISSN Online 2278-1021ISSN Print 2319-5940Since 2012
IJARCCE adheres to the suggestive parameters outlined by the University Grants Commission (UGC) for peer-reviewed journals, upholding high standards of research quality, ethical publishing, and academic excellence.
← Back to VOLUME 15, ISSUE 11, NOVEMBER 2026

Hallucination Detection Using Evidence-Based Methodologies in Large Language Models: Comparative Analysis of Classical vs. Quantum Classifiers

Khushi, Paavai Anand G, Sridevi S, Beenarani Manoj

๐Ÿ‘ 12 views๐Ÿ“ฅ 0 downloads
Share: ๐• f in โœˆ โœ‰
Abstract: Large language models can produce answers that sound right but are not actually supported by the text they were given. This work builds a simple, fully CPU-based check for that problem. Each answer is turned into five numbers: the share of hedge words, the average token likelihood under a small local model, how similar the answer is to the evidence, how similar it is to the question, and an entailment probability. Those five numbers then go into ordinary classifiers (logistic regression, random forest, XGBoost, a tiny neural net) and also into two quantum-style classifiers that run in simulation: a five-qubit variational circuit and a quantum-kernel SVM. We took the first sixty questions from the HaluEval QA set. That gives one hundred and twenty answersโ€”one correct and one hallucinated for every question. Splits keep whole questions together so nothing leaks. On the sixty-answer training slice that both families see, the classical models reach an AUROC of about 0.975. The quantum-kernel SVM gets 0.90 and the variational circuit only 0.53. The test set itself is small (eighteen answers). A later check of the exact quantum state also showed that our particular circuit almost completely ignores its fifth input. The code is reproducible, models reload cleanly, and a small local interface lets anyone try new examples. The pipeline works, but the data set is tiny, calibration is uncertain, and the variational circuit has a clear design flaw, so we make no claim of quantum advantage.

Keywords: hallucination detection, evidence checking, quantum machine learning, quantum kernels, variational quantum classifier, probability calibration

How to Cite:

[1] Khushi, Paavai Anand G, Sridevi S, Beenarani Manoj, โ€œHallucination Detection Using Evidence-Based Methodologies in Large Language Models: Comparative Analysis of Classical vs. Quantum Classifiers,โ€ International Journal of Advanced Research in Computer and Communication Engineering (IJARCCE), DOI: 10.17148/IJARCCE.2026.151173

Creative Commons License This work is licensed under a Creative Commons Attribution 4.0 International License.