Intervention Complexity Proposed as Canonical Reward and Measure of Intelligence

A new academic paper published on arXiv introduces 'intervention complexity' as a novel measure to address limitations in the Legg-Hutter universal intelligence framework. The traditional model relies on externally specified reward functions, which can be arbitrary. This study proposes intervention complexity, characterized by five natural properties including environment-derivedness and universality, serving as a canonical reward indexed by resource bias such as program length or energy. The authors further define intelligence through two dimensions: agent competence relative to an oracle optimum and learning efficiency. A key finding is a separation theorem demonstrating that while action-count intervention complexity is computable in polynomial time, program-length complexity without oracle access remains uncomputable. This gap quantifies the information-theoretic content of learning. The research provides a principled completion of the universal intelligence framework without requiring external normative input, offering significant implications for the development of superintelligence and the pre-training of universal artificial agents.