Omega-S: A Functional Resilience Index for LLM Fine-Tuning

📄 arXiv: 2608.03887v1 📥 PDF

作者: Alberto Acedo

分类: cs.LG, cs.NE, q-bio.MN

发布日期: 2026-08-04

备注: 15 pages of main text plus appendices; 12 tables. Code, per-seed data and all negative results at https://github.com/BiomeMakers/OmegaS-LLM

💡 一句话要点

提出Omega-S以解决大语言模型微调中的能力退化问题

🎯 匹配领域: 支柱九：具身大模型 (Embodied Foundation Models)

关键词: 大语言模型 微调 能力保留 正则化 Omega-S 自然语言处理 机器学习

📋 核心要点

  1. 现有的大语言模型微调方法在新数据上训练时，往往会导致模型原有能力的退化，缺乏有效的保留机制。
  2. Omega-S通过计算权重矩阵的惩罚项来解决这一问题，设计上无需依赖旧数据或复杂的矩阵计算，简化了微调过程。
  3. 在Llama-3-8B模型的实验中，Omega-S在保留原有能力方面表现优异，显著提高了微调后的模型性能。

📝 摘要（中文）

在对大型语言模型进行微调时，模型原有的学习能力会受到影响。本文提出了Omega-S，这是一种仅基于权重矩阵计算的惩罚项，无需先前任务的数据、Fisher矩阵或旧权重的存储。Omega-S的实现仅需在现有训练循环中增加三行代码，且对每步的成本增加不足4%。实验结果表明，在Llama-3-8B模型上，使用Omega-S的微调在保留原有能力方面优于不使用正则化的方法，且在多项对比实验中表现出色。

🔬 方法详解

问题定义：本文旨在解决大型语言模型在微调过程中能力退化的问题。现有方法通常依赖于旧任务的数据或复杂的计算，导致实施困难和效果不佳。

核心思路：Omega-S的核心思路是通过仅使用权重矩阵计算惩罚项，来有效保留模型的原有能力。该方法设计简洁，能够在不依赖旧数据的情况下进行微调。

技术框架：Omega-S的实现非常简单，只需在现有训练循环中添加三行代码。其目标函数基于Tr(A^3)构建，主要通过对节点度的方差进行惩罚来实现。

关键创新：Omega-S的创新之处在于其不依赖于旧权重或复杂的矩阵计算，且通过对节点度的方差进行惩罚，显著提高了模型的能力保留率。这与传统的正则化方法形成了鲜明对比。

关键设计：在实现过程中，Omega-S的设计考虑了多个因素，包括对节点度的方差进行惩罚，确保在不同模块中保持方向一致性。实验中还发现，某些设计选择会导致保留效果的下降。

🖼️ 关键图片

fig_0
img_1
img_2

📊 实验亮点

在Llama-3-8B模型的实验中，使用Omega-S的微调方法在10个种子上保留了62.9%的原有能力，提升至84.1%，且在与调优的权重衰减和EWC方法的对比中，均表现出显著优势，p值分别为0.002和0.014。

🎯 应用场景

Omega-S的研究成果具有广泛的应用潜力，尤其在需要频繁更新和微调的大型语言模型中，如自然语言处理、对话系统和代码生成等领域。通过有效保留模型原有能力，能够提高模型在新任务上的适应性和性能，进而推动相关技术的发展。

📄 摘要（原文）

Fine-tuning a large language model on new data degrades what it previously learned. We present Omega-S, a drop-in penalty computed from the weight matrix alone: it needs no previous-task data, no Fisher matrix and no stored copy of the old weights. It is three lines in an existing training loop and adds under 4% to the cost of a step. Retention. On Llama-3-8B with LoRA, fine-tuned from code to prose and measured by HumanEval over ten seeds, Omega-S retains more of the original capability than no regularisation on 9 of 10 seeds (0.173 -> 0.238 absolute pass@1; sign test one-sided p=0.011, Wilcoxon p=0.006), as a retention ratio, 62.9% -> 84.1%. It also beats tuned weight decay on 10 of 10 seeds (p=0.002) and tuned EWC on 8 of 10 (p=0.014), every arm re-measured in the same session. Mechanism, measured rather than asserted. Omega-S is topological by construction, its objective built from Tr(A^3), but we measured which of its four factors actually moves and three do not: their elasticity with respect to the weights is at or below 1e-4, against 9e-3 for the degree-variance term. As implemented, the composite reduces to a penalty on the variance of node degrees, which means row magnitude in square modules and directional alignment in non-square ones. We report this because a method whose name promises one thing and whose gradient does another should say so. We also enumerate the open design choices, including a contrast-preserving construction that does what it was designed to do and makes retention worse on all ten seeds. Repeating an identical configuration, same seed and same hardware, gives a standard deviation of 0.104 in retention ratio. We have not found this quantified for low-rank fine-tuning of language models, and it bounds every seed-paired comparison in this literature, ours included. Code, per-seed results and the full record of negative results are available.