K. Liu

Yuqiao Tan, Shizhu He, Kang Liu et al.

Large Language Models (LLMs) offer a transparent brain with accessible parameters that encode extensive knowledge, which can be analyzed, located and transferred. Consequently, a key research challenge is to transcend traditional knowledge transfer paradigms rooted in symbolic language and achieve genuine Parametric Knowledge Transfer (PKT). Significantly, exploring effective methods for transferring knowledge across LLMs of different scales through parameters presents an intriguing and valuable research direction. In this paper, we first demonstrate $\textbf{Alignment}$ in parametric space is the fundamental prerequisite to achieve successful cross-scale PKT. We redefine the previously explored knowledge transfer as Post-Align PKT (PostPKT), which utilizes extracted parameters for LoRA initialization and requires subsequent fine-tune for alignment. Hence, to reduce cost for further fine-tuning, we introduce a novel Pre-Align PKT (PrePKT) paradigm and propose a solution called $\textbf{LaTen}$ ($\textbf{L}$oc$\textbf{a}$te-$\textbf{T}$h$\textbf{e}$n-Alig$\textbf{n}$) that aligns the parametric spaces of LLMs across scales only using several training steps without following training. Comprehensive experiments on four benchmarks demonstrate that both PostPKT and PrePKT face challenges in achieving consistently stable transfer. Through in-depth analysis, we identify $\textbf{Neural Incompatibility}$ as the ethological and parametric structural differences between LLMs of varying scales, presenting fundamental challenges to achieving effective PKT. These findings provide fresh insights into the parametric architectures of LLMs and highlight promising directions for future research on efficient PKT. Our code is available at https://github.com/Trae1ounG/Neural_Incompatibility.

Furong Peng, Jinzhen Gao, Xuan Lu et al.

Graph Convolutional Networks (GCNs) suffer from severe performance degradation in deep architectures due to over-smoothing. While existing studies primarily attribute the over-smoothing to repeated applications of graph Laplacian operators, our empirical analysis reveals a critical yet overlooked factor: trainable linear transformations in GCNs significantly exacerbate feature collapse, even at moderate depths (e.g., 8 layers). In contrast, Simplified Graph Convolution (SGC), which removes these transformations, maintains stable feature diversity up to 32 layers, highlighting linear transformations' dual role in facilitating expressive power and inducing over-smoothing. However, completely removing linear transformations weakens the model's expressive capacity. To address this trade-off, we propose Layer-wise Gradual Training (LGT), a novel training strategy that progressively builds deep GCNs while preserving their expressiveness. LGT integrates three complementary components: (1) layer-wise training to stabilize optimization from shallow to deep layers, (2) low-rank adaptation to fine-tune shallow layers and accelerate training, and (3) identity initialization to ensure smooth integration of new layers and accelerate convergence. Extensive experiments on benchmark datasets demonstrate that LGT achieves state-of-the-art performance on vanilla GCN, significantly improving accuracy even in 32-layer settings. Moreover, as a training method, LGT can be seamlessly combined with existing methods such as PairNorm and ContraNorm, further enhancing their performance in deeper networks. LGT offers a general, architecture-agnostic training framework for scalable deep GCNs. The code is available at [https://github.com/jfklasdfj/LGT_GCN].