.
Ground Delay Programs (GDPs) are widely used to mitigate demand-capacity imbalances in air traffic operations. However, existing studies typically address GDP planning from either the perspective of air traffic control or that of individual airlines, with limited attention to their coordination. This study develops a collaborative framework for the airline-driven GDP that integrates proactive slot reduction and airline precancellation. The problem is formulated as a two-stage stochastic program. The first determines slot reductions and flight precancellations in anticipation of various potential capacity scenarios, whereas the second adjusts slot allocations and executes airline recovery actions based on the realized slot capacity. To address the computational challenges of this large-scale problem, we propose an efficient Lagrangian dual decomposition algorithm, which reduces the computation time by 49.06% on average compared with the standard column generation approach. Numerical experiments based on real-world data show that the proposed framework consistently achieves lower expected total recovery costs than rule-based deterministic sequential strategies commonly used in current GDP practice. Moreover, when fairness constraints are imposed, allowing slot exchanges among airlines further reduces total costs by an average of 1.37%. These findings highlight the importance of collaborative decision-making between air traffic managers and airlines, providing a robust and practical tool to enhance system-wide efficiency under capacity disruptions.
Streaming EEG fatigue monitoring is challenging because fatigue signatures are both frequency-dependent and spatially distributed across electrodes, while deployment requires causal updates with bounded latency and memory. Existing Transformers weakly exploit electrode geometry and incur history-dependent cost, whereas graph models typically rely on fixed or unconstrained connectivity that is difficult to adapt as cognitive state changes. We propose HASTE, a hybrid-attention streaming topology-aware estimator that separates where interactions are allowed from how strongly they adapt: a physical k-nearest-neighbor (kNN) graph constrains spatial attention to anatomically feasible neighborhoods, and an online band-aware functional prior biases within-neighborhood attention through an additive log-bias. Causal spectral fusion tracks time-varying band relevance, and causal local temporal attention with a fixed look-back window enables constant-memory inference with respect to session length. Under leave-one-subject-out evaluation with causal normalization on SEED-VIG, SADT, and MPD-DF, HASTE achieves Pearson correlation coefficients (PCCs) of 0.776, 0.705, and 0.669, respectively, outperforming the six baselines evaluated across all three datasets. On SEED-VIG, HASTE also outperforms the added Mamba-DE baseline on all three metrics. On an NVIDIA RTX 3090, its neural forward pass on a precomputed window feature takes ∼2.8ms and uses ∼47MB of graphics processing unit (GPU) memory.
Current benchmarks for Large Language Models (LLMs) focus mainly on classification tasks or require manual evaluation, limiting the accurate assessment of generative capabilities. Inspired by voting games, we propose VotingBench, an automated evaluation framework that measures LLMs' generative performance through default automated peer voting in interactive games, covering both text and multimodal tasks. We position these outputs as scalable and reproducible preference signals rather than human-annotated ground truth. The framework reduces reliance on manual scoring and optionally allows human participation for comparative analysis. Extensive experiments across multiple domains and languages demonstrate strong alignment with existing benchmarks. Human-model comparisons reveal systematic differences in decision-making strategies, highlighting VotingBench as an objective and scalable evaluation tool.
The Chain-of-Thought (CoT) paradigm, while enhancing the interpretability of Large Language Models (LLMs), is constrained by the inefficiencies and expressive limits of natural language. Latent Chain-of-Thought (latent CoT) reasoning, which operates in a continuous latent space, offers a promising alternative but faces challenges from structural complexities in existing multi-step or multi-model paradigms, such as error propagation and coordination overhead. In this paper, we introduce One-Model One-Step, a novel compression framework for Latent Reasoning with Rule-Based Priors(RuPLaR) to address this challenge. Our method trains an LLM to autonomously generate latent reasoning tokens in a single training stage, guided by rule-based prior probability distributions, thereby eliminating cascaded processes and inter-model dependencies. To ensure reasoning quality, we design a joint training objective that enforces answer consistency via cross-entropy, aligns soft tokens with rule-based priors via KL divergence (the Soft Thinking constraint), and adds a problem-thought semantic alignment constraint in the representation space. Extensive experiments show that our compression framework not only improves accuracy by 11.1
This paper investigates the interval estimation problem for a 3-DOF manipulator subject to stealthy attacks. To optimize system resource utilization, a hybrid dynamic event-triggered mechanism is developed. A distinctive feature of this mechanism is the guarantee of a minimum inter-event time, which significantly improves its feasibility for deployment on physical platforms. In the design of the interval observer, measurement noise, nonlinearities, and external disturbances are explicitly considered. By leveraging a coordinate transformation strategy, we construct a framework for robust state estimation. Furthermore, the ultimate boundedness of the state error dynamics is theoretically guaranteed, ensuring stability while effectively alleviating the network communication burden. Finally, the validity of the proposed scheme is demonstrated through a manipulator example.
