InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models

2025-05-20

AI Summary

Existing large language model (LLM) fusion research predominantly focuses on supervised fine-tuning, neglecting the preference alignment (PA) stage; meanwhile, mainstream PA fusion methods (e.g., WRPO) rely solely on source model responses while discarding critical probability distributions. This work pioneers implicit model fusion within the PA framework, introducing a sequence-level multi-model probability fusion mechanism that replaces the single reference model in DPO—thereby avoiding vocabulary alignment challenges and preserving distributional knowledge. We further propose probability clipping and maximum-margin fusion strategies, enabling joint optimization of knowledge distillation and human preference alignment without explicit parameter merging or vocabulary alignment. Evaluated on 11 major benchmarks, our method significantly outperforms prior approaches: using Phi-4 as the pivot model, average score improves from 79.95 to 83.33, with consistent gains across mathematical reasoning, code generation, and general reasoning capabilities.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Humans and AI: Learning Human Values and Preferences

Application Category

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

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.

Problem

• Fusing multiple LLMs via preference alignment unexplored
• Existing PA fusion methods discard probability information
• InfiFPO optimizes fusion with multi-source probability synthesis

Innovation

• Implicit model fusion via preference optimization
• Sequence-level multi-source probability synthesis
• Probability clipping and max-margin fusion strategies