Pack of LLMs: Model Fusion at Test-Time via Perplexity Optimization

TL;DR

This paper introduces PackLLM, a test-time LLM fusion method that optimizes model importance weights based on perplexity to improve task performance, outperforming existing fusion techniques.

Contribution

The paper proposes a novel test-time fusion approach for LLMs using perplexity-based importance weighting, enabling effective integration of arbitrary models during inference.

Findings

Perplexity effectively indicates LLM expertise.

PackLLM outperforms baseline fusion methods by 1.89% accuracy.

Leverages new LLMs to significantly boost performance.

Abstract

Fusing knowledge from multiple Large Language Models (LLMs) can combine their diverse strengths to achieve improved performance on a given task. However, current fusion approaches either rely on learning-based fusers that do not generalize to new LLMs, or do not take into account how well each LLM understands the input. In this work, we study LLM fusion at test-time, which enables leveraging knowledge from arbitrary user-specified LLMs during inference. We introduce Pack of LLMs (PackLLM), an effective method for test-time fusion that leverages each LLM's expertise, given an input prompt. PackLLM performs model fusion by solving an optimization problem for determining each LLM's importance, so that perplexity over the input prompt is minimized. First, our simple PackLLM-sim variant validates that perplexity is a good indicator for measuring each LLM's expertise. Second, our PackLLM-opt variant approximately solves the perplexity minimization problem via a greedy algorithm. The derived importance weights are used to combine the LLMs during inference. We conduct experiments with over 100 total LLMs on a diverse set of tasks. Experimental results show that (i) perplexity is a reliable measure for LLM fusion, (ii) PackLLM outperforms test-time fusion baselines by 1.89% accuracy points, and (iii) PackLLM can leverage new LLMs to improve performance over learning-based fusion approaches by 3.92-11.94% accuracy points.