Mediator: Memory-efficient LLM Merging with Less Parameter Conflicts and Uncertainty Based Routing

TL;DR

Mediator introduces a memory-efficient approach to merge LLMs by layer-wise conflict analysis and uncertainty-based expert routing, achieving better performance with lower costs.

Contribution

The paper proposes a novel layer conflict-aware merging method combined with task uncertainty routing and sparse expert decoupling for efficient LLM model merging.

Findings

Significant performance improvements over existing methods.

Reduced storage and compute costs.

Effective handling of out-of-distribution samples.

Abstract

Model merging aggregates Large Language Models (LLMs) finetuned on different tasks into a stronger one. However, parameter conflicts between models leads to performance degradation in averaging. While model routing addresses this issue by selecting individual models during inference, it imposes excessive storage and compute costs, and fails to leverage the common knowledge from different models. In this work, we observe that different layers exhibit varying levels of parameter conflicts. Building on this insight, we average layers with minimal parameter conflicts and use a novel task-level expert routing for layers with significant conflicts. To further reduce storage costs, inspired by task arithmetic sparsity, we decouple multiple fine-tuned experts into a dense expert and several sparse experts. Considering the out-of-distribution samples, we select and merge appropriate experts based on the task uncertainty of the input data. We conduct extensive experiments on both LLaMA and Qwen with varying parameter scales, and evaluate on real-world reasoning tasks. Results demonstrate that our method consistently achieves significant performance improvements while requiring less system cost compared to existing methods.