POME: Post Optimization Model Edit via Muon-style Projection

2025-10-08

AI Summary

To address the reliance of large language model (LLM) fine-tuning on additional data or computational overhead for performance gains, this paper proposes POME: a general, zero-cost post-fine-tuning optimization method that operates solely on the weight delta ΔW between pre-trained and fine-tuned models. POME innovatively integrates Muon-style projection with truncated singular value decomposition (SVD) to perform low-rank reconstruction of ΔW, explicitly isolating dominant update directions while suppressing noise from small singular values. Crucially, it requires no extra training data, backpropagation, or hyperparameter tuning, and is compatible with diverse fine-tuning paradigms—including RLHF—as well as models ranging from 7B to 72B parameters. Empirical evaluation shows consistent improvements: +2.5% average accuracy on GSM8K and +1.0% on code generation tasks. With its simplicity, stability, and plug-and-play deployment, POME establishes a novel lightweight pre-inference enhancement paradigm for LLMs.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search
• Natural Language Processing: (Large) Language Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

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.

Problem

• Enhancing fine-tuned language models using pretrained and fine-tuned checkpoints
• Applying muon-style projection to weight differences via truncated SVD
• Providing zero-cost performance boost without extra data or optimization

Innovation

• Applies muon-style projection to weight differences
• Uses truncated SVD to equalize dominant update directions
• Provides zero-cost post-processing enhancement for fine-tuning