LLMs Enable Top-Down Knowledge Search for Heuristic Design
A recent paper on arXiv (2605.06123) presents a novel top-down approach for automatic heuristic design (AHD) utilizing large language models (LLMs). In contrast to traditional bottom-up methods that explore executable programs and rely on execution feedback, this strategy prioritizes knowledge as the main search target, with code merely functioning to implement and evaluate it. The authors articulate this transition through a statistical-learning perspective, highlighting a trade-off between distortion and compression, and apply it within both population-based and tree-based AHD frameworks. This knowledge-centric search enhances performance in combinatorial optimization and various other tasks.
Key facts
- arXiv paper 2605.06123 proposes a top-down paradigm for AHD with LLMs
- Knowledge becomes the primary search object, code instantiates and tests it
- Formalized through a statistical-learning view with distortion-compression trade-off
- Instantiated in population-based and tree-based AHD frameworks
- Knowledge-first search improves performance across CO and other tasks
Entities
Institutions
- arXiv