Prompt engineering provides an efficient way to adapt large language models (LLMs) to downstream tasks without retraining model parameters. However, designing effective prompts can be challenging, especially when model gradients are unavailable and human expertise is required. Existing automated methods based on gradient optimization or heuristic search exhibit inherent limitations under black box or limited-query conditions. We propose Domain-Aware Reinforcement Learning for Prompt Optimization (DA-RLPO), which treats prompt editing as a sequential decision process and leverages structured domain knowledge to constrain candidate edits. Our experimental results show that DA-RLPO achieves higher accuracy than baselines on text classification tasks and maintains robust performance with limited API calls, while also demonstrating effectiveness on text-to-image and reasoning tasks.

The growing complexity of real-world optimization problems often necessitates the simultaneous optimization of multiple tasks. Evolutionary multitasking (EMT) algorithms, including the multifactorial evolutionary algorithm (MFEA), have gained significant attention for their ability to solve multiple related optimization tasks concurrently by transferring knowledge across tasks. However, existing EMT methods, particularly MFEA, face challenges in managing dynamic task relationships and ensuring beneficial knowledge transfer, especially in large-scale or highly heterogeneous multitask optimization problems. This paper proposes a novel method, evolutionary multitask optimization with online knowledge transfer and probabilistic outlier detection (OKTPO-MFEA), which incorporates a dynamic adjustment of the random mating probability and a probability-based outlier detection mechanism. The proposed algorithm adapts the knowledge transfer process in real-time, promoting positive knowledge transfer and suppressing negative transfer. In addition, a robust outlier detection strategy is introduced to identify and eliminate harmful genetic contributions, further enhancing optimization performance. The effectiveness of OKTPO-MFEA is validated through extensive experiments on multi-task benchmark problems, demonstrating its superiority in solution quality and convergence speed compared to state-of-the-art methods.

Mild cognitive impairment (MCI), a transitional stage between normal aging and Alzheimer’s disease (AD), comprises three potential trajectories: reversion, stability, or progression. Accurate prediction of these trajectories is crucial for disease modeling and early intervention. We propose a novel analytical framework that integrates a healthy control–AD difference template (HAD) with a large-scale Granger causality algorithm based on long short-term memory networks (LSTM-lsGC) to construct effective connectivity (EC) networks. By applying principal component analysis for dimensionality reduction, modeling dynamic sequences with LSTM, and estimating EC matrices through Granger causality, the framework captures both symmetrical and asymmetrical connectivity, providing a refined characterization of the network alterations underlying MCI progression and reversion. Leveraging graph-theoretical features, our method achieved an MCI subtype classification accuracy of 84.92% (AUC = 0.84) across three subgroups and 90.86% when distinguishing rMCI from pMCI. Moreover, key brain regions, including the precentral gyrus, hippocampus, and cerebellum, were identified as being associated with MCI progression. Overall, by developing a symmetry-aware effective connectivity framework that simultaneously investigates both MCI progression and reversion, this study bridges a critical gap and offers a promising tool for early detection and dynamic disease characterization.