通过权重差值投影提升微调模型性能，无需额外训练。
POME: Post Optimization Model Edit via Muon-style Projection
- 对微调前后权重差进行截断SVD投影，均衡重要更新方向。
- 在GSM8K上提升2.5%，代码生成提升1.0%，效果稳定。
- 零成本适配任意模型与优化器，适合所有微调流程使用。
我们提出后优化模型编辑（POME），一种仅需预训练和微调模型检查点即可提升大语言模型性能的新算法，无需额外数据或进一步优化。核心思想是对微调前后权重差ΔW执行类μ子投影，利用截断奇异值分解（SVD）均衡主导更新方向的影响，并剔除小奇异值（常代表噪声）。作为简单的后处理步骤，POME与训练流程完全解耦，无需修改、无额外开销，兼容任意优化器与分布式框架。POME实现一致增益，在GSM8K上平均性能提升+2.5%，代码生成任务提升+1.0%。其广泛适用性覆盖7B基础模型至72B RLHF指令模型，成为微调流水线中实用的零成本增强方案。代码已公开于https://github.com/NUS-HPC-AI-Lab/POME。
原文摘要 · Abstract (English)
We introduce Post-Optimization Model Edit (POME), a new algorithm that enhances the performance of fine-tuned large language models using only their pretrained and fine-tuned checkpoints, without requiring extra data or further optimization. The core idea is to apply a muon-style projection to $ΔW$, the difference between the fine-tuned and pretrained weights. This projection uses truncated singular value decomposition (SVD) to equalize the influence of dominant update directions and prune small singular values, which often represent noise. As a simple post-processing step, POME is completely decoupled from the training pipeline. It requires zero modifications and imposes no overhead, making it universally compatible with any optimizer or distributed framework. POME delivers consistent gains, boosting average performance by +2.5\% on GSM8K and +1.0\% on code generation. Its broad applicability -- from 7B foundation models to 72B RLHF-instructed models -- establishes it as a practical, zero-cost enhancement for any fine-tuning pipeline. Code is available at https://github.com/NUS-HPC-AI-Lab/POME.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。