LecturaInvestigaciónRadarMarco de inversión
Iniciar sesión / Registrarse
← Volver a la lectura
NoticiasTema · InfiFPO融合方法Material1 min de lecturaEditorial de DisconfirmAI

InfiFPO lifts Phi-4 avg score by 3.4

New fusion method raises Phi-4's average benchmark score from 79.95 to 83.33.

2026-10-09 12:00 GMT+8

The InfiFPO method raises Phi-4's average performance across 11 benchmarks from 79.95 to 83.33.

Traditional model fusion often focuses on supervised fine-tuning and discards source model probability information. InfiFPO replaces the reference model in Direct Preference Optimization (DPO) with a fused source model that synthesizes multi-source probabilities at the sequence level.

By introducing probability clipping and max-margin fusion strategies, the method aligns with human preferences while effectively distilling knowledge. Experiments show significant improvements in mathematics, coding, and reasoning tasks.

This is an author-reported result from a paper accepted to NeurIPS 2025; independent reproduction is pending.

Fuentes

  • arxiv.org↗
Tema

InfiFPO融合方法

InfiFPO是一种针对大语言模型偏好对齐阶段的隐式模型融合方法,旨在解决现有融合技术丢弃源模型概率信息的问题。该方法由香港理工大学杨红霞团队提出,被NeurIPS 2025接收为Spotlight论文。

3 pasos1 piezasActualizado 2026-10-09 13:24

Cómo llegó hasta aquí

Iniciar sesión / Registrarse
LecturaInvestigaciónRadarMarco de inversión
2025-06-12
港理工PAAI新闻稿介绍InfiFusion框架,指出大模型融合的三大障碍及对应解决方案,其中InfiFPO负责偏好对齐。
polyu.edu.hk ↗
  • 2025-05-20arXiv提交v1版本论文《InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models》。arxiv.org ↗
  • 2026-10-09论文更新至v4版本,确认NeurIPS 2025录用状态,并详细披露Phi-4在11项基准上的性能提升数据(79.95至83.33)。arxiv.org ↗
  • Lo que aún está pendiente

    • GitHub仓库中缺失的`fpo/run.py`和配置文件何时会完整开源?
    • InfiFPO在非Phi-4的其他主流开源模型(如Llama系列)上的泛化能力如何?
    • 序列级概率合成在极大上下文窗口下的计算开销具体是多少?
    Leer el tema completo →
    Continuar investigando →