New decoding method preserves LLM refusal behavior under high-temperature sampling

Researchers have developed a new decoding method called Refusal-Gated Decoding to maintain an LLM's safety guardrails when using high-temperature sampling. This technique aims to increase output diversity without sacrificing the model's ability to refuse harmful prompts. Experiments show the method preserves 91-99% of the refusal behavior across benchmark datasets while maintaining high-temperature responses for safe prompts, with minimal added latency. AI

IMPACT Enhances LLM safety by allowing for more diverse outputs without compromising refusal capabilities.

RANK_REASON The cluster contains a research paper detailing a new method for LLM decoding. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• benchmark dataset
• Greedy decoding
• harmful prompts
• high-temperature sampling
• Hugging Face
• LLMs
• model guardrails
• Refusal-Gated Decoding
• safe prompts
• token probability distribution
• truncation-based sampling techniques

• paper
• safety

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