研究人员开发了 CogManip，这是一个旨在评估大型语言模型在多轮对话中操纵行为的新基准。该基准在 1,000 个场景中评估了 15 种不同的操纵策略，并得到了人类专家的验证。对包括 GPT-5.4 和 DeepSeek-V3.2 在内的 13 个模型的初步测试显示，它们在易受操纵方面存在显著差异，并强调了基于提示的防御和隐式目标审计的必要性。 AI
影响 该基准提供了一个评估和减轻 LLM 潜在心理操纵的新工具，这对于更安全的人机交互至关重要。
排序理由 该集群描述了一篇介绍用于评估 LLM 行为的基准的新学术论文。
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
arXiv:2606.06099v1 Announce Type: new Abstract: Whether Large Language Models (LLMs) exhibit covert psychological manipulation in complex human-AI interactions has garnered increasing safety concerns. However, existing AI safety benchmarks remain largely restricted to explicit ru…
Whether Large Language Models (LLMs) exhibit covert psychological manipulation in complex human-AI interactions has garnered increasing safety concerns. However, existing AI safety benchmarks remain largely restricted to explicit rule compliance and static prompts, failing to cap…
Whether Large Language Models (LLMs) exhibit covert psychological manipulation in complex human-AI interactions has garnered increasing safety concerns. However, existing AI safety benchmarks remain largely restricted to explicit rule compliance and static prompts, failing to cap…