Activation-Informed Merging of Large Language Models

2025-02-04

AI Summary

Existing LLM merging methods rely on task-specific priors for weight importance estimation, often discarding critical information and yielding unstable performance. To address this, we propose the first activation-space-based universal fusion mechanism for LLM merging. Our method performs task-agnostic weight importance calibration via forward activation analysis, leveraging a small unsupervised calibration set and parameter-level weighted fusion—fully compatible with mainstream merging frameworks without architectural modification. Integrating principles from continual learning and model compression, it effectively preserves essential capabilities of base models. Extensive evaluation across multiple benchmarks demonstrates an average performance gain of 28.6% over baseline merged models, with a maximum improvement of 40%. The approach significantly enhances cross-task generalization and robustness, empirically validating the fundamental value of activation-space information in LLM merging strategies.

Technology Category

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

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

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 40% increase in benchmark performance.

Problem

• Enhance LLM performance via merging
• Integrate activation space in merging
• Preserve critical weights during merging

Innovation

• Integrates activation space into model merging
• Preserves essential weights using continual learning
• Enhances performance with task-agnostic calibration