This research introduces a model-aware approach to identifying specific text spans where large language models produce hallucinations across 14 languages.

• Develops specialized techniques for detecting and highlighting exact segments of hallucinated content
• Provides a nuanced understanding of how different LLMs produce various types of hallucinations
• Creates a multilingual framework applicable across diverse language contexts
• Contributes to establishing reliability benchmarks for LLM outputs

Security Impact: By precisely identifying AI hallucinations, this research helps protect against misinformation, enhances trust in AI systems, and enables better filtering of unreliable AI-generated content - critical for enterprise deployment of LLM technologies.

HausaNLP at SemEval-2025 Task 3: Towards a Fine-Grained Model-Aware Hallucination Detection