Self-Revising Discovery Systems for Science: A Categorical Framework for Agentic Artificial Intelligence

Researchers present a category-theoretic framework for agentic AI systems that can revise their own representational structures during scientific discovery, rather than merely generating answers within fixed assumptions. The work demonstrates how self-revising discovery systems can be engineered for materials science through two instantiated systems: Builder/Breaker and CategoryScienceClaw.

This arXiv paper addresses a fundamental challenge in AI-driven scientific discovery: how systems can move beyond answer generation to fundamentally restructure the conceptual frameworks in which they operate. The authors employ category theory—a high-level mathematical abstraction—to formalize what happens when discovery systems transition between different "regimes" or representational schemas. Rather than treating the system's foundational assumptions as fixed, they model discovery as verified transitions between categorical frameworks, where old artifacts are preserved and compared against new theoretical structures.

The work emerges from growing recognition that many scientific problems require not just optimization within existing paradigms, but paradigm shifts themselves. Traditional AI systems operate within fixed feature spaces and evaluation criteria; this framework enables systems to question and revise those very foundations. The mathematical rigor provided by category theory offers both theoretical grounding and practical engineering specifications.

For the AI and scientific computing sectors, this represents progress toward more autonomous discovery systems that could accelerate research cycles in materials science, drug discovery, and physics. The two case studies—protein mechanics and fiber-network modeling—demonstrate the framework's feasibility on concrete materials science problems. However, the work remains largely theoretical and academic; widespread practical adoption would require significant additional engineering and validation.

The broader implication centers on developing AI systems capable of genuine conceptual innovation rather than pattern matching. This could eventually impact how discovery is conducted across scientific fields, though commercialization pathways remain unclear. The framework's complexity may limit near-term adoption outside specialized research institutions.

• →Category theory formalizes how AI systems can transition between different representational regimes during scientific discovery.
• →The framework separates retrieval, search, and true discovery through mathematical verification of regime transitions.
• →Prototype systems (Builder/Breaker and CategoryScienceClaw) demonstrate the approach's feasibility in materials science applications.
• →Self-revising discovery systems could accelerate research cycles by enabling paradigm shifts rather than just optimization.
• →The work remains primarily theoretical with limited near-term commercial applications outside specialized research environments.