Wentao Zhang

Jiaxi Cui, Wentao Zhang, Jing Tang et al. · tsinghua

The pervasive deployment of Large Language Models-LLMs in various sectors often neglects the nuanced requirements of individuals and small organizations, who benefit more from models precisely tailored to their specific business contexts rather than those with broadly superior general capabilities. This work introduces \textbf{AnyTaskTune}, a novel fine-tuning methodology coined as \textbf{Task-Fine-Tune}, specifically developed to elevate model performance on a diverse array of domain-specific tasks. This method involves a meticulous process to identify and define targeted sub-tasks within a domain, followed by the creation of specialized enhancement datasets for fine-tuning, thereby optimizing task-specific model performance. We conducted comprehensive fine-tuning experiments not only in the legal domain for tasks such as keyword extraction and sentence prediction but across over twenty different sub-tasks derived from the domains of finance, healthcare, law, psychology, consumer services, and human resources. To substantiate our approach and facilitate community engagement, we will open-source these bilingual task datasets. Our findings demonstrate that models fine-tuned using the \textbf{Task-Fine-Tune} methodology not only achieve superior performance on these specific tasks but also significantly outperform models with higher general capabilities in their respective domains. Our work is publicly available at \url{https://github.com/PandaVT/DataTager}.

Zheng Liu, Mengjie Liu, Siwei Wen et al.

Reinforcement Learning has emerged as the fundamental technique for enhancing reasoning in LLMs. However, existing algorithms apply uniform optimization to all tokens, ignoring their different roles in reasoning process. To address this limitation, we introduce Heterogeneous Adaptive Policy Optimization (HAPO), a comprehensive token-aware algorithm that dynamically adapts optimization based on token entropy. For rollout sampling, we propose Adaptive Temperature Sampling, which adjusts sampling temperature in real time, promoting exploration at high-entropy tokens while preserving coherence at low-entropy ones. For advantage calculation, we introduce Token Level Group Average that normalizes advantages at token level, jointly accounting for sequence-length as in token-mean loss while preserving non-biased treatment. We then develop Differential Advantage Redistribution that leverages entropy and importance ratios to modulate rewards-adjusting updates for tokens with clear signals. For clipping loss, we design Asymmetric Adaptive Clipping, allowing aggressive probability reduction for noisy low-entropy tokens while enabling exploration for high-entropy tokens. Through systematic investigation between entropy and training dynamics, we embedded token-level treatment into every stages to achieve fine-grained control. Extensive experiments demonstrate that HAPO consistently outperforms DAPO across multiple model scales. Our code can be found in https://github.com/starriver030515/HAPO.

Xunkai Li, Zhengyu Wu, Wentao Zhang et al.

Federated Graph Learning (FGL) is a distributed machine learning paradigm that enables collaborative training on large-scale subgraphs across multiple local systems. Existing FGL studies fall into two categories: (i) FGL Optimization, which improves multi-client training in existing machine learning models; (ii) FGL Model, which enhances performance with complex local models and multi-client interactions. However, most FGL optimization strategies are designed specifically for the computer vision domain and ignore graph structure, presenting dissatisfied performance and slow convergence. Meanwhile, complex local model architectures in FGL Models studies lack scalability for handling large-scale subgraphs and have deployment limitations. To address these issues, we propose Federated Graph Topology-aware Aggregation (FedGTA), a personalized optimization strategy that optimizes through topology-aware local smoothing confidence and mixed neighbor features. During experiments, we deploy FedGTA in 12 multi-scale real-world datasets with the Louvain and Metis split. This allows us to evaluate the performance and robustness of FedGTA across a range of scenarios. Extensive experiments demonstrate that FedGTA achieves state-of-the-art performance while exhibiting high scalability and efficiency. The experiment includes ogbn-papers100M, the most representative large-scale graph database so that we can verify the applicability of our method to large-scale graph learning. To the best of our knowledge, our study is the first to bridge large-scale graph learning with FGL using this optimization strategy, contributing to the development of efficient and scalable FGL methods.

Xunkai Li, Zhengyu Wu, Wentao Zhang et al.

Recently, Federated Graph Learning (FGL) has attracted significant attention as a distributed framework based on graph neural networks, primarily due to its capability to break data silos. Existing FGL studies employ community split on the homophilous global graph by default to simulate federated semi-supervised node classification settings. Such a strategy assumes the consistency of topology between the multi-client subgraphs and the global graph, where connected nodes are highly likely to possess similar feature distributions and the same label. However, in real-world implementations, the varying perspectives of local data engineering result in various subgraph topologies, posing unique heterogeneity challenges in FGL. Unlike the well-known label Non-independent identical distribution (Non-iid) problems in federated learning, FGL heterogeneity essentially reveals the topological divergence among multiple clients, namely homophily or heterophily. To simulate and handle this unique challenge, we introduce the concept of structure Non-iid split and then present a new paradigm called \underline{Ada}ptive \underline{F}ederated \underline{G}raph \underline{L}earning (AdaFGL), a decoupled two-step personalized approach. To begin with, AdaFGL employs standard multi-client federated collaborative training to acquire the federated knowledge extractor by aggregating uploaded models in the final round at the server. Then, each client conducts personalized training based on the local subgraph and the federated knowledge extractor. Extensive experiments on the 12 graph benchmark datasets validate the superior performance of AdaFGL over state-of-the-art baselines. Specifically, in terms of test accuracy, our proposed AdaFGL outperforms baselines by significant margins of 3.24\% and 5.57\% on community split and structure Non-iid split, respectively.