New Paper Proposes 'Entity Fidelity' as a General Criterion for Artificial General Intelligence
A recent preprint on arXiv, titled 'On the Definition of Intelligence,' introduces a novel formal standard for artificial general intelligence (AGI) that transcends species and can be utilized across diverse intelligent behavior models. The authors define 'entity fidelity' as the fundamental concept: intelligence entails the capacity to create new entities that embody a concept derived from existing entities. This is articulated as ε-concept intelligence, where an agent is deemed ε-intelligent if no acceptable distinguisher can differentiate between generated and original entities beyond the ε tolerance. The document presents a formal framework, protocols for empirical testing, and explores implications for assessment, safety, and generalization. This preprint can be found on arXiv with the identifier 2507.22423, under the category of Computer Science > Artificial Intelligence.
Key facts
- The paper is titled 'On the Definition of Intelligence'.
- It is available on arXiv with identifier 2507.22423.
- The paper proposes 'entity fidelity' as a general criterion for intelligence.
- Intelligence is defined as the ability to generate entities exemplifying a concept given examples.
- The formalization is called ε-concept intelligence.
- It covers reinforcement learning, generative models, classification, analogical reasoning, and goal-directed decision-making.
- The paper discusses implications for evaluation, safety, and generalization.
- The preprint is categorized under Computer Science > Artificial Intelligence.
Entities
Institutions
- arXiv
- Semantic Scholar