Weighted-Reward Preference Optimization for Implicit Model Fusion

While fusing heterogeneous open-source LLMs with varying architectures and sizes can potentially integrate the strengths of different models, existing fusion methods face significant challenges, such as vocabulary alignment and merging distribution matrices. These procedures are not only complex but also prone to introducing noise and errors. In this paper, we propose an implicit fusion method, Weighted-Reward Preference Optimization (WRPO), which leverages preference optimization between the source LLMs and the target LLM to transfer their capabilities effectively. WRPO eliminates the need for vocabulary alignment and matrix fusion and can be efficiently scaled to accommodate various LLMs. To address distributional deviations between the source and target LLMs, WRPO introduces a progressive adaptation strategy that gradually shifts reliance on preferred examples from the target LLM to the source LLMs. Extensive experiments on the MT-Bench, AlpacaEval-2, and Arena-Hard benchmarks demonstrate that WRPO consistently outperforms existing knowledge fusion methods and various fine-tuning baselines. When applied to LLaMA3-8B-Instruct as the target model, WRPO achieves a length-controlled win rate of 55.9% against GPT-4-Preview-1106 on AlpacaEval-2 and a win rate of 46.2% against GPT-4-0314 on Arena-Hard. Our code is available at https://github.com/SLIT-AI/WRPO. Combining the strengths of multiple Large Language Models (LLMs) can potentially enhance the capabilities of individual models. Model ensemble techniques (Jiang et al., 2023b; Wang et al., 2024b) aggregate predictions from several models to improve overall performance and robustness over a single model. However, this approach requires substantial computational resources, as all models must remain active during inference. The Mixture of Experts (MoE) (Komatsuzaki et al., 2023; Feng et al., 2024; Sukhbaatar et al., 2024) leverages sparse expert networks to boost capacity by activating only a subset of parameters. Despite reduced activation, MoEs still incur significant memory overhead, as all parameters must be maintained. Model merging (Wortsman et al., 2022; Matena & Raffel, 2022; Yadav et al., 2024), which combines independently trained instances of the same model through arithmetic operations, allows a single model to be maintained during inference.