LLM-Based Algorithm Design via Code Graph Corrections
A team of researchers has introduced a cost-effective method for designing automatic algorithms utilizing LLMs, depicting algorithms as directed acyclic graphs. Rather than producing complete algorithms, the system seeks compact corrections from LLMs—operators that can add, substitute, or eliminate code segments. Each correction enhances the graph, allowing for the creation of new algorithms that integrate previous corrections. This strategy breaks down algorithms into correction-level credits, preventing unnecessary rewriting of repeated substructures and retaining important attributes from lower-fitness candidates. The framework aims to optimize realized fitness while keeping computational expenses minimal.
Key facts
- arXiv:2605.10598v1
- LLMs used for automatic algorithm design
- Directed acyclic graph representation of algorithms
- Corrections are compact operators: add, replace, remove code blocks
- Each correction augments the graph
- New algorithms compose with prior corrections
- Correction-level credit assignment
- Maximizes realized fitness under limited computational cost
Entities
Institutions
- arXiv