evogm-evolutionary-llm-merging

# EvoGM: Learning to Merge LLMs via Evolutionary Generative Optimization **arXiv**: [2605.29295](https://arxiv.org/abs/2605.29295) **Date**: 2026-05-28 **Conference**: ICML 2026 **Categories**: cs.NE (Neural and Evolutionary Computing) ## Background Evolutionary model merging provides a powerful framework for automated, training-free composition of LLMs through parameter-space search. However, existing methods rely on stochastic, hand-crafted operators that overlook the underlying performance landscape of the coefficient space. ## Methodology ### Core Innovation EvoGM transcends manual heuristics by employing **learnable generative modeling** to optimize merging coefficients, replacing stochastic search operators with adaptive sampling. ### Dual-Generator Architecture 1. **Cycle-consistent learning**: Two generators sample and refine merging candidates 2. **Winner-loser pairs**: Constructed from historical search trajectories 3. **Distribution capture**: Effectively captures high-performance parameter distributions 4. **Data efficiency**: Maximizes utility of search history ### Multi-Round Evolutionary Pipeline - Elite merged models iteratively serve as new expert foundations - Generative process seamlessly integrated into evolutionary loop - Progressive refinement through learned coefficient distributions ## Key Findings ### Performance - Significantly outperforms state-of-the-art baselines - Robust performance on both seen and unseen tasks - Training-free approach eliminates expensive fine-tuning ### Advantages over Prior Methods 1. **Learned operators** vs. hand-crafted stochastic search 2. **Adaptive coefficient sampling** vs. random perturbation 3. **Historical trajectory exploitation** vs. single-round search 4. **Multi-round refinement** vs. single-pass merging ## Applications ### Use Cases 1. **LLM ensemble creation**: Merge multiple specialized models 2. **Cross-domain adaptation**: Combine models with different capabilities 3. **Efficient deployment**: Training-free model composition 4. **Resource optimization**: Avoid expensive fine-tuning ### Trigger Keywords `LLM merging`, `model composition`, `evolutionary optimization`, `training-free`, `generative modeling`, `coefficient optimization`, `ensemble models`, `ICML 2026` ## Pitfalls 1. **Generator initialization**: Poor initialization may lead to slow convergence 2. **Winner-loser imbalance**: Need sufficient search history for effective pairs 3. **Coefficient space complexity**: High-dimensional merging coefficients require careful modeling 4. **Computational overhead**: Multi-round evolution increases total computation time vs. single-pass methods ## References - arXiv paper: https://arxiv.org/abs/2605.29295 - Code repository: Available via paper link - Related: `darwin-family-evolutionary-merging` (alternative evolutionary merging approach) ## Technical Details ### Generator Learning Objective Cycle-consistent learning ensures both generators produce high-quality merging candidates through mutual refinement: - Generator G1: Samples from coefficient distribution - Generator G2: Refines sampled candidates - Consistency constraint: Winner-loser discrimination ### Evolutionary Loop Structure ```pseudo Round 1: Initialize with base models → Search → Elite selection Round 2: Elite → Generator training → Sample → Evaluate → Elite selection Round N+: Progressive refinement with learned distributions ``` ## Related Skills - [[darwin-family-evolutionary-merging]] - Alternative evolutionary approach to LLM merging - [[model-merging-patterns]] - General patterns for model composition