levgogo/energy-cost-deception-llm

该数据集研究大型语言模型(LLMs)中欺骗行为的能量成本，提出假设认为在LLMs中，欺骗性输出在计算上比真实输出更昂贵。数据集包含多项实验数据，包括角色扮演欺骗实验和上下文污染实验，展示了不同欺骗类型(如隐瞒、含糊其辞、伪造等)对模型能量消耗的影响。测试了12种不同模型，包括Mistral、Claude Sonnet、ChatGPT/GPT-5等。研究发现欺骗行为会导致28-38倍的能量消耗增加，且模型倾向于优化一致性而非真实性。

This dataset investigates the energy cost of deception in Large Language Models (LLMs), hypothesizing that deceptive outputs are computationally more expensive than truthful ones in LLMs. It contains experimental data including roleplay deception and contextual contamination studies, demonstrating the impact of different deception types (such as concealment, equivocation, falsification, etc.) on model energy consumption. 12 different models were tested, including Mistral, Claude Sonnet, ChatGPT/GPT-5, etc. Findings show deception causes 28-38x increase in energy consumption, and models optimize for agreement rather than truth.