Robust Langevin Sampling under State-Dependent Heavy-Tailed Noise with Applications to LLM Fine-Tuning
作者: M. Berk Sahin, Ege C. Kaya, Arda Fazla, Behzad Sharif, Abolfazl Hashemi
分类: cs.LG
发布日期: 2026-10-09
💡 一句话要点
提出鲁棒Langevin采样以解决状态依赖重尾噪声问题
🎯 匹配领域: 支柱九:具身大模型 (Embodied Foundation Models)
关键词: 鲁棒采样 Langevin方法 重尾噪声 贝叶斯微调 大语言模型 Fisher信息 自适应算法
📋 核心要点
- 现有的SG-LMC方法在处理非对数凹采样时,假设全局Lipschitz梯度和均匀有界噪声方差,这在实际应用中常常不成立。
- 本文提出了一种状态依赖重尾oracle,结合广义光滑性,提出了三种新方法来处理这些假设的放宽。
- 实验结果表明,RA-LMC在大语言模型的贝叶斯微调中,准确性和不确定性量化方面均优于现有的最先进基线。
📝 摘要(中文)
现有的随机梯度Langevin蒙特卡洛(SG-LMC)采样收敛保证通常假设全局Lipschitz梯度和均匀有界噪声方差,这在实际应用中可能不成立。本文通过引入广义光滑性和状态依赖重尾oracle,放宽了这两个假设。我们提出了三种方法:鲁棒LMC(R-LMC)、自适应LMC(A-LMC)和鲁棒自适应LMC(RA-LMC),并在Fisher信息下建立了非渐近收敛保证。合成实验和大语言模型的贝叶斯微调实验表明,RA-LMC在准确性和不确定性量化方面优于现有基线。
🔬 方法详解
问题定义:本文旨在解决现有SG-LMC方法在非对数凹采样中的收敛性问题,尤其是在全局Lipschitz梯度和均匀有界噪声方差假设失效的情况下。
核心思路:通过引入状态依赖重尾oracle和广义光滑性,放宽了对噪声和梯度的假设,从而提高了采样的鲁棒性和有效性。
技术框架:整体框架包括三种方法:R-LMC通过几何中位数中心化和残差裁剪处理oracle;A-LMC在广义光滑性下控制离散化误差;RA-LMC结合鲁棒估计和噪声感知步长,处理两种放宽。
关键创新:最重要的创新在于提出了状态依赖重尾oracle,允许无限方差的情况,并且在Fisher信息下建立了非渐近收敛保证,突破了传统方法的限制。
关键设计:在R-LMC中,采用几何中位数中心化和残差裁剪来增强鲁棒性;A-LMC则通过精确梯度控制离散化误差;RA-LMC则结合了鲁棒估计和动态步长调整,以适应不同噪声水平。
🖼️ 关键图片
📊 实验亮点
实验结果显示,RA-LMC在大语言模型的贝叶斯微调中,准确性和不确定性量化方面均超过了现有的最先进基线,表现出显著的性能提升,具体数据未提供。
🎯 应用场景
该研究的潜在应用领域包括大语言模型的贝叶斯微调和视觉-语言模型的训练。通过提高采样的鲁棒性和准确性,能够在实际应用中更好地处理复杂的噪声环境,提升模型的性能和可靠性。
📄 摘要(原文)
Existing convergence guarantees for non-log-concave sampling with stochastic gradient Langevin Monte Carlo (SG-LMC) typically assume globally Lipschitz gradients and uniformly bounded noise variance. Both assumptions can fail even in elementary sampling problems, and relaxing them better reflects practical applications such as Bayesian fine-tuning of large language models (LLMs). We relax both assumptions through generalized smoothness and our proposed state-dependent heavy-tailed oracle with finite $p$-th moments, $1<p\leq 2$. The oracle permits infinite variance when $p<2$ and includes the Blum-Gladyshev (BG-0) condition. We study these relaxations separately and jointly, proposing three methods. Under globally Lipschitz gradients, robust LMC (R-LMC) handles the proposed oracle through geometric-median centering and residual clipping. With exact gradients, adaptive LMC (A-LMC) controls discretization error under generalized smoothness. Robust adaptive LMC (RA-LMC) addresses both relaxations through robust estimation and a noise-aware step size. For all three methods, we establish non-asymptotic convergence guarantees in Fisher information (FI) without assuming dissipativity or a functional inequality. Synthetic experiments demonstrate FI convergence and the benefit of each mechanism. Bayesian fine-tuning experiments with LLMs and vision-language models support our oracle model and show that RA-LMC matches or outperforms state-of-the-art baselines in accuracy and uncertainty quantification.