Decentralized Federated Graph Learning (DFGL) overcomes the potential bottlenecks of the parameter server in FGL. However, extensive cross-worker communication of graph node embeddings during DFGL training introduces substantial communication costs. To improve communication efficiency, constructing sparse network topologies or applying graph sampling are potential methods. In this paper, we first reveal the bidirectional coupling between network topology construction and graph sampling, underscoring the necessity of their joint optimization. Motivated by this insight, we proposeDuplex, a unified framework that co-optimizes these two components by explicitly modeling their interdependent relationship, thereby significantly reducing communication costs while enhancing training performance in DFGL.Duplexformulates the decision-making process as a coordinated configuration$\langle \mathbf {A}, \mathbf {R} \rangle$, where$\bf {A}$is the adjacency matrix of the network topology and$\bf {R}$denotes the set of graph sampling ratios for workers. However, determining proper coordinated configurations to achieve optimal communication efficiency and training performance (e.g., model accuracy and convergence rate) is challenging due to several practical issues,e.g., statistical heterogeneity and dynamic network conditions. To overcome these challenges,Duplexintroduces a novel learning-driven algorithm to adaptively determine optimal network topologies and graph sampling ratios for workers. Experimental results demonstrate thatDuplexreduces completion time by 20.1%–48.8% and communication costs by 16.7%–37.6% to achieve target accuracy, while improving accuracy by 3.3%–7.9% under identical resource budgets compared to baselines.

Large language models (LLMs) rely heavily on sampling methods to generate diverse and highquality text.While existing sampling methods like top-p and min-p have identified the detrimental effects of low-probability tails in LLMs' outputs, they still fail to effectively distinguish between diversity and noise.This limitation stems from their reliance on probability-based metrics that are inherently sensitive to temperature scaling.Through empirical and theoretical analysis, we make two key discoveries: (1) the pre-softmax logits exhibit a clear statistical separation between informative tokens and noise, and (2) we prove the mathematical equivalence of min-p and top-(1-p) under uniform distribution over logits.These findings motivate the design of top-nσ, a novel sampling method that identifies informative tokens by eliminating noise directly in logit space.Unlike existing methods that become unstable at high temperatures, top-nσ achieves temperature-invariant token selection while preserving output diversity.Extensive experiments across reasoning and creative writing tasks demonstrate that our method consistently outperforms existing approaches, with particularly significant improvements in high-temperature settings.