A HYBRID MULTI-SIGNALFRAMEWORK FOR HALLUCINATION DETECTION IN LARGE LANGUAGE MODELS

Abstract

Large Language Models (LLMs) such as GPT-4, LLaMA, and Gemini have demonstrated strong capabilities in generating fluent and contextually relevant responses across diverse natural language tasks. However, such models suffer from hallucinations, generating in formation that is erroneous or deceptive and may decrease the veracity of the model, es pecially for critical applications like medicine, law, and science. Prevailing tech- niques for hallucination detection have tended to rely on just one form of verification, such as re trieval grounding, semantic consistency, and confidence estimation, all of which are sub ject to shortcomings when applied individually. The thesis develops a novel Hybrid Hallucination Detection Framework that uses multiple verification signals for the purpose of factual verification. The novel framework uses retrieval similarity with the use of TF-IDF and cosine similarity, semantic similarity by means of Sentence-BERT (SBERT) similarity, and confidence scoring using uncertainty measures to verify the factual accuracy of generated claims based on evidence retrieved from Wikipedia. Furthermore, an Adaptive Weight Optimizer dynamically adjusts the contribution of each signal to improve robustness across varying claim types. The effectiveness of the proposed framework is evaluated on a controlled benchmark dataset and publicly available FEVER and VitaminC datasets. Experimental observations demonstrate effective separation between factual and hallucinated claims, highlighting the capability of the hybrid scoring mechanism to improve hallucination detection reliability. The proposed approach offers a lightweight and practical framework for enhancing factual consistency in LLM-generated responses without requiring extensive model retraining or fine-tuning.