Evolutionary Optimization of Model Merging Recipes

2024-03-19

arXiv.org

AI Summary

Existing large language model (LLM) merging techniques suffer from unstable fusion performance, heavy reliance on manual expertise, and high computational overhead. Method: This paper proposes an automated model merging framework based on Differential Evolution (DE), the first to systematically integrate multi-objective evolutionary algorithms into model merging recipe search. It jointly optimizes critical merging strategies—including weight allocation, layer selection, and normalization schemes—to simultaneously enhance performance, efficiency, and generalization. Leveraging a parameterized merging template and a zero-shot task validation mechanism, the framework eliminates the need for additional fine-tuning or labeled data. Contribution/Results: Evaluated across 12 downstream tasks, the method achieves an average accuracy improvement of 2.7%, significantly outperforming handcrafted recipes and simple averaging baselines. Moreover, its search cost is reduced by 90% compared to exhaustive grid search.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search
• Natural Language Processing: (Large) Language Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search

Problem

• Automatic Model Combination
• Resource Efficiency
• Cross-domain Application

Innovation

• Automated Model Fusion
• Evolutionary Process Mimicry
• Cross-Domain Model Creation