MERGEvolve framework blends model merging with evolutionary exploration

A team of researchers (Chao Wang, Yuchen Guo, Zheng Tan, Guanchun Wang, Yanbiao Ma, Qiqi Duan, Peng Wu) has introduced MERGEvolve, a novel framework that fuses model merging with evolutionary strategies. Traditional model merging combines multiple expert models into one by averaging their weights, but this method only explores within the convex combination space — the direct interpolation of parameters. MERGEvolve breaks this limitation by treating the merged model as a starting point for evolutionary optimization. During the merging phase, expert models serve as deterministic sources to create a strong initial point. In the evolution phase, random noise is added to explore arbitrary directions in parameter space, allowing the model to discover high-performance regions that pure merging cannot reach.

The paper provides theoretical analysis showing that MERGEvolve can find solutions outside the convex hull of the experts. Extensive experiments on both single-task and multi-task benchmarks demonstrate that the framework consistently achieves performance competitive with state-of-the-art model merging baselines. Ablation studies confirm that the quality of the initial merged model is critical for efficient evolutionary exploration — a poor start limits the benefits of noise-based search. The work suggests that treating model merging as initialization for evolution is a promising direction for creating powerful multi-task models without additional training data.