Detecting Hallucinations in LLMs: Semantic Consistency Approach

Hallucinations — where LLMs generate confident but incorrect information — remain one of the biggest challenges in deploying AI systems. Our research at IEEE InCODE-2026 proposes a novel approach. ## The Problem LLMs like GPT-4, Claude, and Llama can produce highly convincing but factually wrong outputs. This is particularly dangerous in: - Medical advice - Legal analysis - Financial reporting - Scientific citations ## Our Approach: Semantic Consistency The key insight: if an LLM's output is factually correct, multiple semantically equivalent paraphrases should agree. If the model is hallucinating, paraphrases will show inconsistencies. ## How It Works 1. Generate the model's output 2. Create 3-5 semantic paraphrases 3. Encode each with sentence embeddings 4. Measure pairwise consistency 5. Flag outputs with low consistency scores ## Results Our method achieves 0.89 F1 score across 5 LLMs and 3 benchmarks, a 12% improvement over existing methods. This work has direct implications for building trustworthy AI systems, especially in RAG pipelines where hallucination detection is critical.