EvoGM: Learning to Merge LLMs via Evolutionary Generative Optimization
Southern University of Science and Technology and Pengcheng lab · Southern University of Science and Technology · University of Chinese Academy of Sciences · Pengcheng Laboratory · Peng Cheng Laboratory · Southern University of Science and Technology of China
PDF 由论文原始站点提供，PaperCompass 不保存论文文件。
摘要
Evolutionary model merging provides a powerful framework for the automated, training-free composition of LLMs through parameter-space search. However, existing methods predominantly rely on stochastic, hand-crafted operators that overlook the underlying performance landscape of the coefficient space. We propose Evolutionary Generative Merging (EvoGM), a framework that transcends manual heuristics by employing learnable generative modeling to optimize merging coefficients. Specifically,, EvoGM features a dual-generator architecture with cycle-consistent learning to adaptively sample and refine promising merging candidates. By constructing winner-loser pairs from historical search trajectories, our framework effectively captures high-performance parameter distributions and maximizes data efficiency. This generative process is seamlessly integrated into a multi-round evolutionary pipeline, where elite merged models iteratively serve as new expert foundations. Extensive experiments across diverse benchmarks demonstrate that EvoGM significantly outperforms state-of-the-art baselines, exhibiting robust performance on both seen and unseen tasks.