A Compositional Framework for Open-ended Intelligence

2026-06-13 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

A new compositional framework for open-ended intelligence, published on 2026-06-13, defines this capacity as adapting to novel problems and environments beyond training data. The framework formalizes open-ended intelligence as a closure ℬ(P,C) induced by a finite primitive set P and composition operators C. This closure supports unbounded compositional generation across diverse tasks and worlds. The underlying mathematics requires representational primitives (e.g., states, actions) and algorithmic primitives (e.g., nearest neighbor), combined with composition motifs like recursion and sequencing. This structure enables infinite adaptive responses. The framework supports research into evaluation metrics for interpretability and architectures designed for native compositional generalization. A novel architectural objective, "next primitive prediction," is proposed, aiming to acquire reusable algorithmic primitives and their compositional grammar through recombination. Curriculum learning and self-play are identified as mechanisms for lifelong learning and expanding the framework's closure, grounded by case studies in physics, evolution, and neuroscience.

Key takeaway

For AI Scientists developing adaptive systems, this framework suggests shifting focus towards architectural objectives like "next primitive prediction." You should prioritize acquiring reusable algorithmic primitives and their compositional grammar, rather than just task-specific learning. This approach, supported by curriculum learning and self-play, can enable systems to generate infinite adaptive responses to novel problems, significantly enhancing generalization capabilities beyond current methods.

Key insights

Open-ended intelligence is formalized as a compositional closure of primitives and operators, enabling unbounded adaptive generation.

Principles

• Open-ended intelligence requires compositional grammar.
• Lifelong learning expands primitive sets and motifs.
• Recombination generates novel solutions.

Method

The framework proposes "next primitive prediction" as an architectural objective. This trains models to acquire reusable algorithmic primitives and their compositional grammar, generating new solutions via recombination.

In practice

• Design architectures for compositional generalization.
• Implement curriculum learning for primitive discovery.
• Utilize self-play to expand adaptive responses.

Topics

• Open-ended Intelligence
• Compositional Generalization
• Algorithmic Primitives
• Next Primitive Prediction
• Lifelong Learning
• Curriculum Learning

Best for: AI Scientist, Research Scientist

Related on AIssential

• MOSAIC: Modular Orchestration for Structured Agentic Intelligence and Composition — MOSAIC structures LLM-based agentic model selection through blueprints, improving automated data science performance and traceability.
• Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability — Reasoning SFT can generalize cross-domain, but it is conditional on optimization, data quality, and model capability.
• Learning to Theorize the World from Observation — Understanding the world means inferring explicit, compositional programs from raw observations, not just predicting future states.
• Structural Generalization on SLOG without Hand-Written Rules — A neural cellular automaton can achieve structural generalization in semantic parsing without hand-written rules.
• What happens when AI starts building itself? — Recursive Superintelligence seeks to build truly self-improving AI through open-endedness, automating research and development.
• AEL: Agent Evolving Learning for Open-Ended Environments — AEL enables LLM agents to learn from experience by dynamically selecting memory retrieval policies and integrating reflective causal insights.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.