Abstract
Model fusion combines multiple Large Language Models (LLMs) with different strengths into a more powerful, integrated model through lightweight training methods. Existing works on model fusion focus primarily on supervised fine-tuning (SFT), leaving preference alignment (PA) --a critical phase for enhancing LLM performance--largely unexplored. The current few fusion methods on PA phase, like WRPO, simplify the process by utilizing only response outputs from source models while discarding their probability information. To address this limitation, we propose InfiFPO, a preference optimization method for implicit model fusion. InfiFPO replaces the reference model in Direct Preference Optimization (DPO) with a fused source model that synthesizes multi-source probabilities at the sequence level, circumventing complex vocabulary alignment challenges in previous works and meanwhile maintaining the probability information. By introducing probability clipping and max-margin fusion strategies, InfiFPO enables the pivot model to align with human preferences while effectively distilling knowledge from source models. Comprehensive experiments on 11 widely-used benchmarks demonstrate that InfiFPO consistently outperforms existing model fusion and preference optimization methods. When using Phi-4 as the pivot model, InfiFPO improve its average performance from 79.95 to 83.33 on 11 benchmarks, significantly improving its capabilities in mathematics, coding, and reasoning tasks.
InfiFPO icon and overview.
InfiFPO: A Paradigm Shift in Model Fusion
Our solution, InfiFPO, introduces three key breakthroughs:
InfiFPO methodology overview.
Ⅰ. Implicit Model Fusion🎯
Instead of wrestling with complex vocabulary alignment at the token level, we operate at the sequence level, seamlessly integrating probabilities from multiple source models. This elegant approach preserves crucial probability information while avoiding compatibility headaches.
Ⅱ. Three Pillars of Stability 🔧
-
Length Normalization
- Eliminates bias from varying tokenization patterns across models
- Ensures fair comparison of sequence probabilities regardless of token length
-
Probability Clipping
- Prevents underperforming source models from introducing noise
- Maintains training stability by setting intelligent probability boundaries
-
Max-Margin Fusion
- Automatically identifies the most informative source model for each scenario
- Focuses on learning distinctive, complementary knowledge
Ⅲ. Efficient Training Pipeline ⚡
By transforming the complex reinforcement learning problem into an efficient offline optimization objective, InfiFPO achieves remarkable results without the computational overhead of traditional methods.
Impressive Results Across the Board
Our comprehensive evaluation across 11 diverse benchmarks demonstrates InfiFPO’s consistent superiority:
| Capability Area | Before InfiFPO | After InfiFPO | Improvement |
|---|---|---|---|
| Mathematics | 72.85 | 75.80 | +2.95 |
| Coding | 79.47 | 85.15 | +5.68 |
| Overall | 79.95 | 83.33 | +3.38 |
Beyond the Numbers: Technical Foundation
InfiFPO’s beauty lies in its mathematical foundation. By replacing the reference model in Direct Preference Optimization with a carefully fused source model, we create an optimization objective that simultaneously:
- Aligns with human preferences
- Distills knowledge from multiple expert models
- Maintains training stability and efficiency
The extra gradient analysis reveals how InfiFPO weights training samples based on the divergence between source and pivot model preferences, focusing optimization efforts where they matter most.
BibTeX
@misc{gu-2025-infifpo,
title={InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models},
author={Yanggan Gu and Zhaoyi Yan and Yuanyi Wang and Yiming Zhang and Qi Zhou and Fei Wu and Hongxia Yang},
year={2025},
eprint={2505.13878},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2505.13878},
}