Learnable Novelty Unifies AI Objectives for Intelligence

Yanbo Zhang, Michael Levin· July 22, 2026 View original

Summary

This research proposes "learnable novelty" as a unified objective for intelligence, distinguishing between information a learner can convert into knowledge and unlearnable surprise. It demonstrates how this single differentiable quantity can drive complexity generation, abstraction, and exploration across various AI tasks.

Traditional AI objectives, such as novelty search and the free-energy principle, often fail because they conflate all forms of surprise, whether it's information a system can actually learn from or just random noise. This paper introduces a new concept called "learnable novelty," which specifically targets the portion of information that a learner can convert into actionable knowledge. The researchers developed a closed-form estimator for learnable novelty, built on a computationally efficient and differentiable reservoir computer. This estimator, when used as a measure without any supervision, successfully classifies complexity, ranking Turing-complete systems highest. Furthermore, when applied as an objective function, learnable novelty drives neural cellular automata to generate complex behaviors like solitons, organizes image representations in an unsupervised manner for tasks like MNIST, and significantly enhances exploration in reinforcement learning agents, outperforming baselines in most environments. This work suggests a unified quantitative foundation for diverse aspects of intelligence, including complexity generation, abstraction, and exploration.

Why it matters

This offers a potentially unifying theory for various AI objectives, which could lead to more general and efficient AI systems capable of unsupervised learning, complex behavior generation, and robust exploration.

How to implement this in your domain

  1. 1Explore integrating learnable novelty as an intrinsic reward mechanism in your reinforcement learning environments to improve exploration.
  2. 2Investigate applying the learnable novelty estimator for unsupervised feature learning or representation organization in your data pipelines.
  3. 3Consider using this framework to analyze and classify the complexity of emergent behaviors in your AI systems.
  4. 4Experiment with reservoir computing architectures as a component for estimating learnable novelty in your models.

Who benefits

AI/ML DevelopmentRoboticsGamingScientific ResearchAutonomous Systems

Key takeaways

  • "Learnable novelty" distinguishes useful information from unlearnable surprise.
  • A single differentiable objective can unify complexity generation, abstraction, and exploration.
  • The proposed estimator improves unsupervised learning and reinforcement learning exploration.
  • This concept offers a new quantitative foundation for understanding intelligence.

Original post by Yanbo Zhang, Michael Levin

"arXiv:2607.18433v1 Announce Type: new Abstract: Intelligence appears under different names in different fields: as data compression in statistics and machine learning, as universal computation in dynamical systems, and as adaptive behavior in agents. Each field carries its own ob…"

View on X

Originally posted by Yanbo Zhang, Michael Levin on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI Engineering & DevToolsAI Research

New Tool Generates Contamination-Resistant, Labeled Code Datasets for LLMs

Spaghetti Architect is a new open-source tool that generates controlled, multi-language code datasets, addressing issues of contamination and lack of semantic control in existing code corpora. It creates correct-by-construction programs with adjustable "messiness" and difficulty labels, making it ideal for training and evaluating code-generating LLMs.

Yuxiang JiJul 22, 2026
AI ResearchAI Engineering & DevTools

New Method Safely Gates Hazardous LLM Knowledge Without Deletion

Researchers introduce Token Inoculation, a method that allows large language models to retain sensitive "dual-use" knowledge while selectively refusing hazardous queries. This approach uses a special token to condition the model's behavior, improving safety without sacrificing benign domain performance.

Seunghyun Lee, Dongyoon Han, Sangdoo YunJul 22, 2026
AI ResearchAI Engineering & DevTools

GNNAS-TSP Selects Optimal Algorithms for Traveling Salesman Problem

Researchers introduce GNNAS-TSP, a Graph Neural Network (GNN)-based framework for automated algorithm selection (AS) for the Traveling Salesman Problem (TSP). GNNAS-TSP learns TSP instance representations directly from raw graph data, avoiding manual feature engineering, and formulates AS as a joint cost-prediction and ranking task to select the best solver from a portfolio under fixed computational budgets.

Zhaoxuan Li, Jiale Yang, Yifei Lu, Mustafa MisirJul 22, 2026