Evolving Parallel Algorithm Portfolios via Potential-Aware Instance Generation with LLMs
arXiv:2608.06808v1 Announce Type: new Abstract: The Automatic Construction of Portfolios via Large Language Models (LLM-ACP) suffers from poor generalization in practical few-shot scenarios when solving complex combinatorial optimization problems. Instance and algorithm co-evolution frameworks address this by expanding the training dataset with generated hard instances on which the current algorithm portfolio underperforms, thereby enhancing generalization. However, this paradigm faces two criti
arXiv에서 원문 보기