Recently, large language models (LLMs) have demonstrated superior performance across various tasks by adhering to scaling laws, which significantly increase model size. However, the huge computation overhead during inference hinders the deployment in industrial applications. Many works leverage traditional compression approaches to boost model inference, but these always introduce additional training costs to restore the performance and the pruning results typically show noticeable performance drops compared to the original model when aiming for a specific level of acceleration. To address these issues, we propose a fine-grained token-wise pruning approach for the LLMs, which presents a learnable router to adaptively identify the less important tokens and skip them across model blocks to reduce computational cost during inference. To construct the router efficiently, we present a search-based sparsity scheduler for pruning sparsity allocation, a trainable router combined with our proposed four low-dimensional factors as input and three proposed losses. We conduct extensive experiments across different benchmarks on different LLMs to demonstrate the superiority of our method. Our approach achieves state-of-the-art (SOTA) pruning results, surpassing other existing pruning methods. For instance, our method outperforms BlockPruner and ShortGPT by approximately 10 points on both LLaMA2-7B and Qwen1.5-7B in accuracy retention at comparable token sparsity levels.

Tabular data optimization methods aim to automatically find an optimal feature transformation process that generates high-value features and improves the performance of downstream machine learning tasks. Current frameworks for automated feature transformation rely on iterative sequence generation tasks, optimizing decision strategies through performance feedback from downstream tasks. However, these approaches fail to effectively utilize historical decision-making experiences and overlook potential relationships among generated features, thus limiting the depth of knowledge extraction. Moreover, the granularity of the decision-making process lacks dynamic backtracking capabilities for individual features, leading to insufficient adaptability when encountering inefficient pathways, adversely affecting overall robustness and exploration efficiency. To address the limitations observed in current automatic feature engineering frameworks, we introduce a novel method that utilizes a feature-state transformation graph to effectively preserve the entire feature transformation journey, where each node represents a specific transformation state. During exploration, three cascading agents iteratively select nodes and idea mathematical operations to generate new transformation states. This strategy leverages the inherent properties of the graph structure, allowing for the preservation and reuse of valuable transformations. It also enables backtracking capabilities through graph pruning techniques, which can rectify inefficient transformation paths. To validate the efficacy and flexibility of our approach, we conducted comprehensive experiments and detailed case studies, demonstrating superior performance in diverse scenarios.