Source-linked AI summary
Agentic programs: an emerging form of scientific software in computational materials science
Yunsung Lim, Haekwan Jeon, Jaesun Kim, Jisu Kim, Seungwu Han
TL;DR
Computational materials science has largely left scientific judgment outside computational workflows. This paper proposes agentic programs, which combine deterministic algorithms, bounded LLM judgment, verification, and episodic maturation, and demonstrates the approach with DeMARS for modeling disordered crystals.
Problem
Scientific workflows efficiently automate well-defined computation but still leave many important research decisions to human reasoning.
Method
The paper defines agentic programs as software that combines conventional algorithms with bounded LLM judgment, task-specific verification, episodic maturation, and production-time delegation, illustrated by DeMARS.
Results
Mature DeMARS processed randomly selected batches of disordered CIFs end-to-end without routine case-by-case human involvement, producing certified models or verifiable refusals.
Takeaways & Limitations
Agentic programs can delegate bounded scientific judgment while preserving deterministic checks and explicit failure when evidence or verification criteria are insufficient.
Takeaways & Limitations
LLM-based judgments remain sensitive to context and model, and hallucinations or missing physical and chemical considerations may evade deterministic gates.
Abstract
from arXiv · showhide
Computational materials science has traditionally delegated algorithmic tasks to computers while leaving scientific judgments to humans. We argue that recent LLM-based agent harnesses enable an emerging form of scientific software, agentic programs, that combine deterministic algorithms with bounded LLM-based judgment, task-specific verification, episodic maturation, and complete delegation in production. We illustrate this concept with DeMARS, an agentic program for constructing atomistic models from experimentally measured disordered crystal structures.