Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Yanggan Gu, Yuanyi Wang, Zhaoyi Yan, Yiming Zhang, Qi Zhou, Fei Wu, Hongxia Yang

TL;DR

InfiFPO tackles the challenge of fusion during the preference alignment phase by replacing the DPO reference with a fused source-model distribution that preserves sequence-level probabilities. It derives an offline objective (FuseRLHF) and introduces stability mechanisms—length normalization, probability clipping, and max-margin fusion—to robustly distill diverse source-model knowledge into a pivot. Empirical results across 11 benchmarks show InfiFPO consistently outperforms existing model fusion and preference optimization baselines, achieving notable gains in math and coding capabilities while reducing training time. This approach offers a scalable, information-rich path for integrating heterogeneous LLMs without vocabulary conflicts, enhancing overall alignment and capability.

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: Implicit Model Fusion via Preference Optimization in Large Language Models

TL;DR

InfiFPO tackles the challenge of fusion during the preference alignment phase by replacing the DPO reference with a fused source-model distribution that preserves sequence-level probabilities. It derives an offline objective (FuseRLHF) and introduces stability mechanisms—length normalization, probability clipping, and max-margin fusion—to robustly distill diverse source-model knowledge into a pivot. Empirical results across 11 benchmarks show InfiFPO consistently outperforms existing model fusion and preference optimization baselines, achieving notable gains in math and coding capabilities while reducing training time. This approach offers a scalable, information-rich path for integrating heterogeneous LLMs without vocabulary conflicts, enhancing overall alignment and capability.

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.

Paper Structure

This paper contains 33 sections, 28 equations, 2 figures, 16 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. InfiFPO: Preference Optimization for Model Fusion
  4. FuseRLHF: RLHF for Implicit Model Fusion
  5. InfiFPO: Efficient Preference Optimization for IMF
  6. Gradient Analysis.
  7. Experiments
  8. Setup
  9. Main Results
  10. Analysis
  11. Albation Study
  12. Related Work
  13. Conclusion
  14. Mathematical Derivations
  15. Deriving the FPO Objective
  16. ...and 18 more sections

Figures (2)

  • Figure 1: Overview of InfiFPO for implicit model fusion. We compute probabilities for preferred ($\boldsymbol{y}_w$) and dispreferred ($\boldsymbol{y}_l$) responses using both pivot and source models. Following length normalization and probability clipping, we identify the source model with the maximum normalized probability difference from the pivot model for fusion and preference alignment.
  • Figure 2: Different fusion strategies.