Activation-Informed Merging of Large Language Models

TL;DR

Activation-Informed Merging (AIM) addresses the challenge of fusing multiple fine-tuned LLMs while preserving the base model's capabilities. By analyzing the activation space of a task-agnostic calibration set, AIM derives per-channel saliency and applies an adaptive relaxation to weight updates, making the merging process more robust to outliers and data quality. The method is designed as a plug-in that can augment any existing merging approach, and extensive experiments across Llama-2 and Qwen-VL architectures show consistent improvements, with HV Gain increases up to $40\%$ in some cases. The work demonstrates the practical value of incorporating activation information into merging and suggests directions for further theoretical and architectural enhancements in multi-task LLM fusion.

Abstract

Model merging, a method that combines the parameters and embeddings of multiple fine-tuned large language models (LLMs), offers a promising approach to enhance model performance across various tasks while maintaining computational efficiency. This paper introduces Activation-Informed Merging (AIM), a technique that integrates the information from the activation space of LLMs into the merging process to improve performance and robustness. AIM is designed as a flexible, complementary solution that is applicable to any existing merging method. It aims to preserve critical weights from the base model, drawing on principles from continual learning (CL) and model compression. Utilizing a task-agnostic calibration set, AIM selectively prioritizes essential weights during merging. We empirically demonstrate that AIM significantly enhances the performance of merged models across multiple benchmarks. Our findings suggest that considering the activation-space information can provide substantial advancements in the model merging strategies for LLMs, with up to a 40% increase in benchmark performance.

Figures (4)

• Figure 1: Overview of the proposed activation-informed merging (AIM) in LLMs.
• Figure 2: The Pareto fronts of models under different scenarios. Note that the points in these plots represent all models benchmarked in Table \ref{['tab:results']}, for better readability, we only visualize the dominating points in each case. The measured increases in HV Gain when AIM is applied can be clearly seen in the Pareto frontier shifting further forward when AIM is applied compared to when only a population of merged models is evaluated.
• Figure 3: The Impact of the Relaxation Factor $\omega$ on Merged Model Performance. This figure plots the relative change in HV-Gain compared to scenarios without AIM. The x-axis represents $1-\omega$, reflecting that decreasing $\omega$ results in more relaxation. The plot indicates that for some tasks, smaller values of $\omega$ continue to yield benefits. An $\omega$ of 0.4-0.6 appears to strike a balance.
• Figure 4: AIM's Robustness to Calibration Set Size. HV-Gain is plotted against the number of calibration blocks (log scale) for the DARE TIES merge. The dashed line is the baseline performance without AIM. Significant improvement is achieved with small data, and performance stabilizes at only 8 blocks.