Yucheng Xie, Fu Feng, Jing Wang, Xin Geng, Yong Rui
原文摘要
Generative model pre-training typically optimizes a model as a whole without explicitly organizing its parameters for selective reuse. Decomposable pre-training instead learns reusable components for downstream adaptation while retaining strong generative performance. However, existing knowledge diversion methods incur substantial parameter overhead through weight factorization and rely on fixed component assignments to induce specialization, potentially compromising in-distribution generation quality. We therefore propose Specificity-Guided Knowledge Diversion (SGKDiv), which learns how generative knowledge is shared and specialized for effective generation and selective downstream reuse. SGKDiv constructs Knowledge Units from paired feed-forward channels, preserving native computational structure without SVD-based reparameterization. Calibrated condition-wise routing combines a shared usage prior with conditional adjustments, while complementary regularization shapes their allocation, diversity, and magnitude. Together, these designs replace predefined component assignments with adaptively learned and explicitly regulated knowledge specialization. Experiments across diverse model and training settings demonstrate better in-distribution generation than dense and MoE baselines, alongside stronger downstream adaptation than representative PEFT methods across eight datasets.
Stochastic first-order methods use sampled gradients to reduce per-iteration costs, but estimation errors can slow convergence and, in some settings, cause divergence. To improve gradient accuracy while retaining acceleration, we develop scalar and diagonal versions of VR-Adam-SHANG for smooth stochastic convex optimization. The methods combine STORM-type recursive estimation, Adam-style preconditioning, and SHANG's accelerated interpolation and correction, using at most two stochastic gradient evaluations per iteration without full-gradient refreshes. Under unbiased sampling, bounded variance, and mean-square smoothness of gradient noise, polynomial steps with prescribed initialization yield $O(1/K)$ mean-square gradient-estimation error and expected objective-gap bounds combining $O(1/K^2)$ initial-energy decay with an $O(1/\sqrt K)$ stochastic term after $K$ iterations. These guarantees require neither monotonicity of the preconditioner nor a uniform positive lower bound on its eigenvalues. The analysis also establishes polynomial-step convergence of projected nonrecursive Adam-SHANG-s, filling a gap in its original theory. Experiments compare the methods at matched sample-gradient budgets. In every tested setting, at least one of the two versions achieves the lowest mean terminal objective gap among all compared methods.