Improving instruction hierarchy in frontier LLMs

【RSS 原文 · 未经 AI 筛选或翻译】IH-Challenge trains models to prioritize trusted instructions, improving instruction hierarchy, safety steerability, and resistance to prompt injection attacks.