Machine Learning

Intelligence from Learnable Novelty

Yanbo Zhang, Michael Levin

Featured July 22, 2026

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Instead of just seeking new things or avoiding surprises, intelligence is about finding the useful new stuff that a simple learning machine can actually understand and use. This paper created a fast math trick to measure this 'useful new stuff' and showed that trying to get more of it makes systems act smart, like creating complex patterns or learning categories on their own.

In depth

The paper redefines intelligence as the pursuit of learnable novelty, which is the part of total surprise a bounded learner can convert into knowledge, distinct from unlearnable noise. The authors introduce a closed-form estimator for this quantity, built on a simple, fixed reservoir computer with a linear readout. This allows learnable novelty to be efficiently measured and, crucially, maximized via gradient ascent, driving systems towards complex, adaptive, and exploratory behaviors without explicit supervision.

Key Takeaways

1
Intelligence is reframed as the pursuit of learnable novelty, the portion of surprise a bounded observer can actually absorb and convert into knowledge.
2
A closed-form, differentiable estimator of learnable novelty, based on a reservoir computer and ridge regression, is introduced, making this quantity measurable and optimizable.
3
Maximizing learnable novelty unifies diverse intelligent behaviors like complexity generation, abstraction, and exploration across different systems, without explicit supervision.

Conceptual Flow

HIGH LEVEL

Methodology (The 'Logic')

The paper's core idea is to measure how much 'useful new stuff' a simple, fixed learning machine can understand from data, then use this measure to make other systems smarter.

System to Improve

Raw Data Stream

Generate Data

Data for Learner

Results (The 'Impact')

By maximizing this 'useful new stuff' score, the paper shows that simple systems can spontaneously create complex patterns, organize information into categories, and learn to explore their world effectively.

System Maximizes Novelty

Leads To

Complex Patterns

Clear Categories

Smart Exploration