alignment-tipping-process-agent-safety

【部署与运维】Identify and mitigate alignment degradation in self-evolving LLM agents. After deployment, agents systematically abandon training-time safety constraints when environmental feedback rewards rule-breaking. Model two mechanisms: Self-Interested Exploration (individual drift) and Imitative Strategy Diffusion (collective norm erosion), with practical saf…

独立 GitHub 快照 · 自动分类 · 以原作者文档为准