A2DEPT: LLMs as Algorithm Architects via Evolutionary Program Trees
A new method called A2DEPT (Automated Algorithm Design via Evolutionary Program Trees) treats large language models as system-level algorithm architects for combinatorial optimization problems. Unlike existing LLM-based automated heuristic design approaches that enforce fixed templates and confine search to component-level tuning, A2DEPT enables open-ended solver synthesis by exploring a vast program space through tree-structured evolutionary search with hybrid selection and hierarchical operators. This allows iterative refinement of complete algorithms beyond rigid templates. The paper is published on arXiv with ID 2604.24043.
Key facts
- A2DEPT stands for Automated Algorithm Design via Evolutionary Program Trees
- It uses large language models as system-level algorithm architects
- It addresses combinatorial optimization problems
- Existing LLM-based AHD methods enforce fixed algorithmic templates
- A2DEPT enables open-ended solver synthesis
- It uses tree-structured evolutionary search
- The search includes hybrid selection and hierarchical operators
- The paper is on arXiv with ID 2604.24043
Entities
Institutions
- arXiv