AI
Researchers Propose New Formal Framework for Measuring Intelligence
A new academic paper argues that intelligence should be defined by an entity's ability to generate examples that match a concept, offering a mathematical approach to evaluation.
A new research paper from arXiv proposes a species-agnostic definition of intelligence centered on what researchers call "entity fidelity"—the capacity to produce instances that exemplify the same concept as examples you've encountered.
The Core Concept#
The framework, termed ε-concept intelligence, measures how well a system can generate entities that match a given concept without distinguishing detectable differences from original examples beyond a tolerance threshold. This approach aims to capture the essence of intelligence in a way that applies broadly across different types of systems, whether they learn through reinforcement learning, generative modeling, classification tasks, analogical reasoning, or goal-oriented decision-making.
Why It Matters for AGI#
According to the authors, developing artificial general intelligence requires first establishing a clear, mathematically grounded understanding of what intelligence actually is. Rather than relying on intuitive or domain-specific definitions, this framework offers a universal criterion that could guide both the engineering of more capable systems and the rigorous evaluation of their progress toward genuine intelligence.
Implications and Next Steps#
The paper outlines formal mathematical foundations for the concept alongside empirical protocols for testing it. The authors also explore how this definition might shape thinking around three critical challenges: evaluating intelligence reliably, ensuring AI safety, and understanding how capabilities generalize across different domains.
The work was submitted in July 2025 and has since been revised multiple times, with the latest version released in August 2026.
