研究人员开发了 RevengeBench，这是一个新的基准测试，旨在根据 AI 代理在游戏环境中观察到的行为来逆向工程其决策程序。该基准测试使用了来自 CodeClash 锦标赛的 75 个 LLM 生成的策略，允许学习者通过创建自定义对手策略来设计受控实验，从而引发目标 AI 的信息性行为。这种方法旨在提高策略的可解释性并实现对手建模，重建的策略显示出可衡量的竞争优势，特别是对于较弱的模型。 AI
影响 能够更好地理解 AI 决策制定并提高对手建模能力。
排序理由 该条目描述了 arXiv 上发布的一个新基准测试和方法，用于从行为实验中逆向工程 AI 策略。[lever_c_demoted from research: ic=1 ai=1.0]
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
arXiv:2606.26094v1 Announce Type: new Abstract: For most of scientific history, researchers studying behavior could only infer hidden mechanisms from outward actions: an inverse problem that becomes more tractable when observation is augmented by targeted intervention. We pose a …
For most of scientific history, researchers studying behavior could only infer hidden mechanisms from outward actions: an inverse problem that becomes more tractable when observation is augmented by targeted intervention. We pose a computational analogue: given only behavioral tr…
For most of scientific history, researchers studying behavior could only infer hidden mechanisms from outward actions: an inverse problem that becomes more tractable when observation is augmented by targeted intervention. We pose a computational analogue: given only behavioral tr…