Ambiguity processing in Large Language Models : Detection, resolution, and the path to hallucination

Ambiguity in Large Language Models (LLMs) serves as a primary pathway to hallucination generation, undermining trust in AI systems across critical applications. This survey provides the first comprehensive analysis linking ambiguity processing to hallucination mechanisms, establishing how unresolved linguistic, knowledge-based, pragmatic, and cross-modal ambiguities create pathways to fabricated information. Through systematic examination of theoretical foundations, detection methodologies, and resolution approaches, we reveal that current LLMs process ambiguity through constraint satisfaction mechanisms, yet exhibit poor correlation between internal representations and human disambiguation strategies. Analysis of thirty specialized benchmarks demonstrates that while sophisticated resolution methods show promise, from strategic prompting to multi-agent collaboration, fundamental limitations persist in computational scalability, cultural adaptation, and real-world deployment. The findings identify four critical research frontiers: developing architectures with native uncertainty representation, bridging scalability gaps between laboratory success and deployment constraints, creating culturally-aware disambiguation strategies, and establishing verification frameworks for genuine understanding. These challenges represent fundamental tests of AI system intelligence that will determine whether future LLMs can navigate linguistic uncertainty effectively or remain constrained by brittle pattern matching.