Delta is an inference-time technique that significantly reduces hallucinations in LLMs without requiring model retraining or additional data.

• Works by randomly masking parts of the input and contrasting different generation paths
• Improves reliability by identifying and eliminating potentially fabricated content
• Operates during inference, making it compatible with existing LLM deployments

Medical Impact: Delta addresses a critical challenge in healthcare applications where factually incorrect AI outputs could lead to harmful clinical decisions, enhancing LLM trustworthiness in sensitive medical contexts.

Delta -- Contrastive Decoding Mitigates Text Hallucinations in Large Language Models