DIESEL introduces a novel semantic guidance mechanism for filtering undesired content from Large Language Models without requiring expensive retraining.

• Uses comparison to reference embeddings to dynamically detect and filter unsafe content
• Achieves high effectiveness against adversarial jailbreaking attacks
• Provides computational efficiency compared to other alignment techniques
• Maintains model performance while enhancing security guardrails

This research addresses critical security concerns in AI deployment by offering a practical approach to prevent LLMs from generating harmful or unaligned responses, essential for responsible AI implementation in business contexts.

DIESEL -- Dynamic Inference-Guidance via Evasion of Semantic Embeddings in LLMs