LLM-Driven Framework Enhances Algorithm Portfolios via Potential-Aware Instance Generation
A new framework, Potential-aware Instance and Algorithm Co-evolution (PIAC), has been introduced to improve the generalization of Automatic Construction of Portfolios via Large Language Models (LLM-ACP) in few-shot scenarios. The framework addresses two critical limitations of existing instance and algorithm co-evolution approaches: the reliance on high-quality reference solutions for evaluating instance hardness, and the limited diversity of generated instances due to single-mode generation patterns. PIAC introduces a novel metric called potential gain, which estimates the generalization gain by perturbing the generated algorithms, thereby eliminating the need for reference solutions. This allows for more efficient and diverse instance generation, enhancing the performance of algorithm portfolios on complex combinatorial optimization problems. The research is detailed in a paper available on arXiv (arXiv:2608.06808), submitted as a new announcement. The work is significant for the field of AI-driven algorithm design, offering a more robust method for training algorithm portfolios in practical settings where reference solutions are often unavailable.
Key facts
- The framework is called Potential-aware Instance and Algorithm Co-evolution (PIAC).
- It targets the Automatic Construction of Portfolios via Large Language Models (LLM-ACP).
- PIAC addresses poor generalization in few-shot scenarios for combinatorial optimization problems.
- It overcomes limitations of existing co-evolution frameworks: need for reference solutions and limited instance diversity.
- The core contribution is a novel metric called potential gain.
- Potential gain estimates generalization gain by perturbing generated algorithms.
- The research is available on arXiv with identifier arXiv:2608.06808.
- The paper was announced as a new submission on arXiv.
Entities
Institutions
- arXiv