A Hybrid Framework for Hallucination Detection in Large Language Models

Hallucination in Large Language Models (LLMs) poses a significant threat to their reliability. However, leveraging LLMs themselves for hallucination detection remains computationally intensive and often yields inconsistent results. This challenge underscores the need for more efficient and specialized models tailored to the task. In response, this paper proposes a decoupled framework for hallucination detection, challenging the prevailing assumption that only large-scale models can address complex safety issues in natural language processing (NLP). This paper introduces a modular framework that separates language representation from task-specific classification. Frozen transformer encoders (BERT, RoBERTa, DeBERTa) serve as universal feature extractors, providing rich linguistic representations, while lightweight neural classifiers (ANN, RNN, BiRNN, LSTM, BiLSTM, GRU) perform the hallucination detection task. This separation yields models with fewer than 2% of the parameters of standard LLMs, while retaining linearO(n × d) complexity, enabling efficient deployment under real-world resource constraints. Experimental results, measured by accuracy, precision, recall, and F1-score, demonstrate that the best models rival ChatGPT-3.5 (175B) in zero-shot settings on PolyFEVER and outperform instruction-tuned models like LLaMA-7B, 13B, and 70B. On the FactCHD benchmark, they consistently surpass larger LLMs, and on HaluEval, they deliver competitive or superior performance across dialogue, QA, summarization, and general tasks. These results highlight the strength and efficiency of encoder–classifier architectures for hallucination detection, making them lightweight yet powerful tools for fact verification, content moderation, and enhancing LLM reliability.