Abstract
<title>Abstract</title> <p> Large language models (LLMs) have achieved remarkable success in natural language understanding and generation, enabling applications across scientific research, healthcare, education, and decision support systems. Despite these advances, LLMs frequently generate fluent but factually incorrect responses, a phenomenon commonly referred to as hallucination. Such behavior raises critical concerns regarding reliability, accountability, and safe deployment of generative AI systems. Existing approaches to mitigate hallucinations including retrieval-augmented generation, multi-agent debate frameworks, and confidence calibration techniques primarily focus on improving response quality but often lack a systematic mechanism for estimating the trustworthiness of model outputs. In particular, many current methods fail to address the problem of overconfident incorrect predictions, where models produce highly confident answers despite limited or inaccurate knowledge, a phenomenon we refer to as the accuracy paradox. In this research paper, we introduce <bold>TrustSLM</bold> , a trust-aware evaluation framework designed to assess and regulate responses generated by multiple large language models. The proposed framework integrates outputs from multiple base models and evaluates them using a set of reliability metrics that capture semantic consensus, response uncertainty, evidence consistency, answerability, manipulation signals, and overconfidence detection. These metrics are aggregated into a trust feature representation that is processed by a lightweight classifier to estimate an overall trust score for each generated response. Based on this trust score, the system determines whether the response should be accepted, hedged with uncertainty, or rejected through a controlled abstention policy. We evaluate the proposed framework using multiple prompt datasets and compare it against baseline approaches including single-model generation and majority voting across models. Experimental results demonstrate that TrustSLM significantly improves hallucination detection and trust calibration while reducing overconfident incorrect responses. The proposed architecture introduces a trust-aware control layer for large language model systems and contributes toward safer, more interpretable, and reliable deployment of generative AI technologies in high-stakes domains. </p>