Online symbolic construction of causal world models.
Continuous learning and repair integrated into the agent’s decision loop.
Application of Meta‑Interpretive Learning with dynamic predicate invention for reusable abstractions.
Generation of a hierarchical, disentangled set of high‑quality concepts.
Scalability to domains with complex relational dynamics where propositional approaches suffer combinatorial explosion.
Achieves sample‑efficiency orders of magnitude higher than the PPO neural‑network baseline.
📖 Full Retelling
WHO: Enrique Crespo‑Fernandez, Oliver Ray, Telmo de Menezes e Silva Filho, and Peter Flach.
WHAT: They propose a framework that builds symbolic causal world models online by integrating continuous learning and repair, using Meta‑Interpretive Learning and dynamic predicate invention to create reusable abstractions and a hierarchy of disentangled concepts.
WHERE: The work was published on the arXiv AI preprint server under cs.AI, identified as arXiv:2602.17217.
WHEN: It was submitted to arXiv on 19 February 2026.
WHY: To address the sample inefficiency, lack of transparency, and poor scalability of conventional world‑modeling methods, enabling agents to internalize the logic of complex environments more efficiently.
No entity connections available yet for this article.
Deep Analysis
Why It Matters
This research offers a new way to build causal world models online, improving transparency and scalability while drastically reducing data needs. It could enable smarter agents that learn more efficiently in complex environments.
Context & Background
Continual learning seeks to adapt models over time
Symbolic causal models provide interpretable reasoning
Meta-Interpretive Learning enables rule induction
Predicate invention creates reusable abstractions
Sample efficiency is a key challenge for neural methods
What Happens Next
Future work will explore deploying the framework in robotics and autonomous systems, integrating it with existing reinforcement learning pipelines, and expanding the approach to larger relational domains.
Frequently Asked Questions
What is dynamic predicate invention?
It is a technique that automatically generates new predicates to capture relationships in data, enabling more expressive symbolic models
How does this method improve sample efficiency?
By building high-level abstractions, the agent needs fewer observations to learn causal dynamics compared to flat neural networks
Will the code be publicly available?
The authors plan to release the implementation on a public repository after the paper is published
How does it compare to PPO?
The approach achieves orders of magnitude higher sample efficiency while maintaining comparable or better performance in complex relational tasks
Original Source
--> Computer Science > Artificial Intelligence arXiv:2602.17217 [Submitted on 19 Feb 2026] Title: Continual learning and refinement of causal models through dynamic predicate invention Authors: Enrique Crespo-Fernandez , Oliver Ray , Telmo de Menezes e Silva Filho , Peter Flach View a PDF of the paper titled Continual learning and refinement of causal models through dynamic predicate invention, by Enrique Crespo-Fernandez and 3 other authors View PDF HTML Abstract: Efficiently navigating complex environments requires agents to internalize the underlying logic of their world, yet standard world modelling methods often struggle with sample inefficiency, lack of transparency, and poor scalability. We propose a framework for constructing symbolic causal world models entirely online by integrating continuous model learning and repair into the agent's decision loop, by leveraging the power of Meta-Interpretive Learning and predicate invention to find semantically meaningful and reusable abstractions, allowing an agent to construct a hierarchy of disentangled, high-quality concepts from its observations. We demonstrate that our lifted inference approach scales to domains with complex relational dynamics, where propositional methods suffer from combinatorial explosion, while achieving sample-efficiency orders of magnitude higher than the established PPO neural-network-based baseline. Subjects: Artificial Intelligence (cs.AI) Cite as: arXiv:2602.17217 [cs.AI] (or arXiv:2602.17217v1 [cs.AI] for this version) https://doi.org/10.48550/arXiv.2602.17217 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Enrique Crespo Fernandez [ view email ] [v1] Thu, 19 Feb 2026 10:08:31 UTC (989 KB) Full-text links: Access Paper: View a PDF of the paper titled Continual learning and refinement of causal models through dynamic predicate invention, by Enrique Crespo-Fernandez and 3 other authors View PDF HTML TeX Source view license Current browse c...
