New TERM defense improves LLM jailbreak robustness without target

Researchers have developed a new defense mechanism called TERM (Target-free Entropy-guided Residual Modulation) to combat jailbreak attacks on large language models. This inference-time defense adapts to various prompts by analyzing attention entropy deviations, calibrating them against a baseline to distinguish harmful from benign requests. TERM then modulates the model's internal residual updates without relying on predefined safety directions, demonstrating improved robustness against jailbreaks while minimizing over-refusal on sensitive but safe prompts across models like Llama, Qwen, and Gemma. AI

IMPACT Enhances LLM safety by providing a novel defense against adversarial prompting, potentially leading to more reliable deployment of AI systems.

RANK_REASON This is a research paper detailing a new method for LLM safety. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Gemma
• Hugging Face
• large language models
• Llama
• Qwen
• Qwen3

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →