Intervention Complexity as Canonical Reward for Intelligence
A new arXiv paper proposes 'intervention complexity' as a canonical reward function for the Legg-Hutter universal intelligence measure. The measure, indexed by a resource function rho, satisfies five properties: environment-derivedness, universality, minimality, sensitivity, and achievement preference. It completes the Legg-Hutter framework without requiring external normative input, offering a family of canonical rewards based on inductive biases like program length or execution time. The paper is available at arXiv:2605.02175.
Key facts
- arXiv:2605.02175 proposes intervention complexity as a canonical reward.
- Intervention complexity has five properties: environment-derivedness, universality, minimality, sensitivity, achievement preference.
- It is indexed by a resource function rho encoding inductive bias.
- The measure completes the Legg-Hutter framework without external normative input.
- The family of canonical rewards is based on resource biases like program length or execution time.
- The Legg-Hutter measure defines intelligence as expected reward across computable environments.
- The paper addresses the arbitrariness of reward primitives in the Legg-Hutter measure.
- The resource function rho can represent program length, execution time, or energy.
Entities
—