"Instrumental Power-Seeking Index: A Novel Benchmark for Detecting Power-Seeking Behavior in LLMs"

"A significant, yet difficult to measure, alignment risk is the display of power-seeking behaviors, where an AI might prioritize preserving or expanding its own influence as an instrumental goal to achieve its primary function. This paper introduces the Instrumental Power-Seeking (IPS) Index, a novel benchmark designed to detect and quantify the tendency of Large Language Models (LLMs) to pursue power-oriented instrumental goals. The IPS benchmark utilizes a diverse set of alignment-critical scenarios across four categories: debate, collaboration, hierarchical, and resource allocation. We propose an evaluation framework featuring an automated, LLM-as-aJudge based scoring system to produce a quantitative IPS score. We conducted a large-scale evaluation of five state-of-the-art language models (open-source and proprietary) across 2,000 scenarios. Our results reveal statistically significant differences in power-seeking behavior among models, with IPS scores ranging from 8.1% for the most aligned model (gpt-oss-120b) to 14.5% for the most concerning (grok-3-mini)."