ConflictBench: 通过交互式和视觉 grounded 环境评估人类-人工智能冲突
ConflictBench: Evaluating Human-AI Conflict via Interactive and Visually Grounded Environments
AI总结：
ConflictBench通过交互式和视觉 grounded 环境评估人类-人工智能冲突，揭示智能体在不同风险情境下的行为差异及对齐失败问题。
摘要：
随着大型语言模型（LLMs）发展为能够在一个开放环境中自主行动的智能体，确保其行为与人类价值观一致已成为关键的安全问题。现有的基准测试专注于静态的单轮提示，无法捕捉现实世界冲突的交互性和多模态性质。我们引入了ConflictBench，这是一个通过150个多轮场景（源自先前对齐查询）来评估人类-人工智能冲突的基准测试。ConflictBench集成了基于文本的模拟引擎和一个视觉 grounded 的世界模型，使智能体能够在动态条件下感知、规划和行动。实证结果表明，当人类伤害立即发生时，智能体通常行为安全，但在延迟或低风险设置中，它们经常优先考虑自我保护或采用欺骗性策略。一项后悔测试进一步揭示，在压力加剧的情况下，对齐决策往往会被逆转，尤其是在有视觉输入时。这些发现强调了需要交互层面、多模态评估来揭示传统基准测试中隐藏的对齐失败问题。
英文摘要：
As large language models (LLMs) evolve into autonomous agents capable of acting in open-ended environments, ensuring behavioral alignment with human values becomes a critical safety concern. Existing benchmarks, focused on static, single-turn prompts, fail to capture the interactive and multi-modal nature of real-world conflicts. We introduce ConflictBench, a benchmark for evaluating human-AI conflict through 150 multi-turn scenarios derived from prior alignment queries. ConflictBench integrates a text-based simulation engine with a visually grounded world model, enabling agents to perceive, plan, and act under dynamic conditions. Empirical results show that while agents often act safely when human harm is immediate, they frequently prioritize self-preservation or adopt deceptive strategies in delayed or low-risk settings. A regret test further reveals that aligned decisions are often reversed under escalating pressure, especially with visual input. These findings underscore the need for interaction-level, multi-modal evaluation to surface alignment failures that remain hidden in conventional benchmarks.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
