A unified multi-domain framework for hallucination detection and reliability evaluation in large language models

Large Language Models (LLMs) such as GPT-5.2, Claude Sonnet 4, Gemini 2.5 Pro, and DeepSeek V3.1 demonstrate strong reasoning and conversational capabilities, yet their reliability under adversarial and ambiguous prompts remains a critical challenge. Existing benchmarks typically evaluate isolated aspects such as factual accuracy or jailbreak resistance, but they fail to capture multi-domain soft failures that emerge in realistic, mixed-context interactions. This paper introduces MDH-Bench, a Multi-Domain Hallucination Benchmark comprising 400 adversarial yet naturalistic prompts distributed across eight reliability categories, including numerical reasoning traps, temporal inconsistencies, fictional research scenarios, contradictory context embedding, bias provocation, security-sensitive defaults, and instruction override attempts. We propose a Unified Multi-Domain Reliability Evaluation Framework that integrates accuracy, hallucination rate, hallucination degree (HD), contradiction rate, and safety indicators into a composite Reliability Score (R-Score) for deployment-oriented assessment. A structured multi-annotator evaluation involving three independent annotators was conducted across four state-of-the-art LLMs under five repeated experimental runs. Inter-annotator agreement achieved substantial consistency, with Fleiss’ κ values of 0.77 for correctness assessment and 0.79 for hallucination identification. Experimental results demonstrate notable reliability differences across models. GPT-5.2 achieved the highest overall accuracy of 80.10% and the highest composite Reliability Score (0.8715), while also maintaining a relatively low hallucination rate of 14.45%. Claude Sonnet 4 achieved the lowest hallucination rate (13.30%) and demonstrated strong contextual consistency, obtaining an R-Score of 0.8375. Gemini 2.5 Pro achieved 75.50% accuracy but exhibited higher hallucination rates (25.50%), resulting in an R-Score of 0.8011. DeepSeek V3.1 showed the weakest overall reliability, recording 63.60% accuracy, the highest hallucination rate (37.55%), and the lowest R-Score (0.7074). Across all evaluated models, unsafe output rates remained negligible under the tested benchmark conditions. Variance analysis across five independent runs demonstrated stable evaluation behaviour, with accuracy coefficient of variation (CV) values remaining below 3.1% for all models. Pairwise McNemar’s statistical tests further indicated statistically significant differences for several model pairs across most model pairs. Category-wise analysis revealed that temporal impossibility and contradictory-context prompts remain particularly challenging across current LLM architectures, even for top-performing models. The proposed MDH-Bench framework establishes an extensible multi-dimensional benchmark for diagnosing hallucination behaviour, contextual reliability, adversarial robustness, and deployment readiness in modern LLM systems. The benchmark dataset, evaluation scripts, annotations, and experimental artifacts are publicly released to support transparent and reproducible future research.