This repository contains model artifacts (LoRA adapter) for
EvoGM, as presented in the paper
EvoGM: Learning to Merge LLMs via Evolutionary Generative Optimization.
EvoGM (Evolutionary Generative Merging) is a framework that addresses evolutionary model merging as a learnable search problem. Instead of relying on hand-crafted mutation or crossover operators, it treats validation performance as feedback and learns where high-quality merging coefficients are likely to lie through learnable generative modeling.
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, the framework effectively captures high-performance parameter distributions.
1@article{jiang2024evogm,
2 title={EvoGM: Learning to Merge LLMs via Evolutionary Generative Optimization},
3 author={Tao Jiang and others},
4 journal={arXiv preprint arXiv:2605.29295},
5 year={2026}
6}
