Model Merging in LLMs, MLLMs, and Beyond: Methods, Theories, Applications and Opportunities

🌈 Abstract

The paper provides a comprehensive overview of model merging methods and their applications in various domains. It proposes a new taxonomy that divides existing model merging methods into two stages: pre-merging and during-merging. The paper also discusses the application of model merging in foundation models, including large language models, multimodal large language models, and image generative models, as well as its use in different machine learning subfields such as continual learning, multi-task learning, domain generalization, federated learning, few-shot learning, and adversarial learning. Finally, the paper highlights the remaining challenges and future research directions in the field of model merging.

🙋 Q&A

[01] Methodology Overview

1. What are the key techniques in the pre-merging stage of model merging? The pre-merging stage aims to create better conditions for merging and includes the following key techniques:

• Linearization fine-tuning to achieve weight space and input space disentanglement
• Architecture transformation to convert heterogeneous models into homogeneous models
• Weight alignment to place models in the same basin

2. What are the main categories of during-merging methods? The during-merging methods focus on designing sophisticated techniques to merge multiple models into one, and can be divided into:

• Basic merging methods that perform simple parameter merging strategies
• Weighted merging methods that merge models based on importance calculated by specific rules
• Subspace merging methods that project models into sparse subspaces for merging
• Routing-based methods that dynamically merge models based on input samples during inference
• Post-calibration based methods that correct the merged model

3. What theoretical or empirical analyses have been conducted on the effectiveness of model merging? Theoretical and empirical analyses of model merging have focused on three main aspects:

1. Model merging of different checkpoints in the same training trajectory, which can be explained by reducing variance and helping the model converge to flat local optima.
2. Model merging of different models fine-tuned on the same dataset, which can be explained by the linear mode connectivity property of neural networks.
3. Model merging of different models fine-tuned on different datasets or tasks, which requires weight disentanglement as a necessary condition for effective merging.

[02] Application Overview

1. How can model merging be applied to large language models (LLMs)? Model merging can be applied to LLMs in the following ways:

• Enhancing the domain-specific capabilities of pre-trained LLMs by merging models fine-tuned on different tasks
• Mitigating untruthfulness and toxicity in LLM outputs by merging models with different reward alignments
• Achieving knowledge unlearning by merging the pre-trained model with a model fine-tuned to forget specific knowledge
• Accelerating the training of LLMs by merging checkpoints during the training process

2. How can model merging be applied to multimodal large language models (MLLMs)? Model merging can be used to:

• Merge models from different modalities (e.g., image, audio, video, language) into a single multimodal model
• Transfer knowledge from high-resource modalities to low-resource modalities through cross-modal knowledge transfer

3. How can model merging be applied in different machine learning subfields? Model merging has been applied in the following subfields:

• Continual learning: To mitigate catastrophic forgetting of old tasks
• Multi-task/multi-domain/multi-objective learning: To facilitate knowledge transfer across tasks, domains, or objectives
• Domain/out-of-distribution generalization: To improve model robustness and generalization to unseen distributions
• Federated learning: To aggregate local models from different clients
• Zero-shot/few-shot learning: To enhance cross-task generalization with limited data
• Adversarial learning: For both attack and defense strategies, as well as copyright protection

[03] Future Directions

1. What are the remaining challenges in the field of model merging? Some key challenges include:

• As the number of tasks increases, the performance gap between existing merging methods and independent expert models becomes larger.
• Current model merging methods incur high memory costs during the merging process.
• Lack of trust guarantees and in-depth theoretical analysis for model merging techniques.

2. What are the future research directions in model merging? Potential future research directions include:

• Developing more efficient and scalable model merging methods to handle a large number of tasks
• Providing theoretical guarantees and analysis for the performance and robustness of merged models
• Exploring cross-disciplinary applications of model merging beyond the machine learning domain
• Investigating the use of model merging for trustworthy and accountable AI systems