LLM Agentic Search Discovers Improved Density Functional
A recent study presents an agentic search system that employs large language models to autonomously identify exchange-correlation functionals for density functional theory. Documented in arXiv:2605.05460, the system suggests structured modifications of functional forms, influenced by evolutionary history, within an iterative loop of planning, execution, and summarization. To assess performance, functional parameters are optimized using a standard thermochemistry dataset and tested on a separate subset. The top-performing functional identified, SAFS26-a (Seed Agentic Functional Search 2026), surpasses the established ωB97M-V benchmark. This research marks a transition from manually crafted functionals to automated discovery, utilizing LLMs to navigate an extensive design landscape.
Key facts
- arXiv:2605.05460 describes an agentic search system for XC functionals.
- The system uses an LLM to propose functional-form changes guided by evolutionary history.
- Improvements are measured by optimizing parameters against a thermochemistry dataset.
- The strongest discovered functional is SAFS26-a.
- SAFS26-a improves upon the ωB97M-V baseline.
- The system operates in an iterative plan-execute-summarize loop.
- Most XC functionals have been hand-designed by humans.
- The approach automates the functional design process.
Entities
—