We investigate the capacity of Large Language Models (LLMs) for imaginative reasoning—the proactive construction, testing, and revision of hypotheses in information-sparse environments. Existing benchmarks, often static or focused on social deduction, fail to capture the dynamic, exploratory nature of this reasoning process. To address this gap, we introduce a comprehensive research framework based on the classic "Turtle Soup" game, integrating a benchmark, an agent, and an evaluation protocol. We present TurtleSoup-Bench, the first large-scale, bilingual, interactive benchmark for imaginative reasoning, comprising 800 turtle soup stories sourced from both the Internet and expert authors. We also propose Mosaic-Agent, a novel agent designed to assess LLMs' performance in this setting. To evaluate reasoning quality, we develop a multi-dimensional protocol measuring logical consistency, detail completion, and conclusion alignment. Experiments with leading LLMs reveal clear capability limits, common failure patterns, and a significant performance gap compared to humans. Our work offers new insights into LLMs' imaginative reasoning and establishes a foundation for future research on exploratory agent behavior.

Anomaly detection is one of the most significant tasks in multivariate time series analysis, while it remains challenging to model complex patterns for improving detection accuracy and to interpret the root causes of anomalies. However, existing studies either consider only the temporal dependencies, or simply reconstruct the original input for detection, both neglecting the hidden relationships among multivariate. We propose an adversarial graph neural network based anomaly detection model, called SGAT-AE, which consists of aSelf-learningGraphATtention network (SGAT), anAuto-Encoder (AE), and an adversarial training component. Specifically, SGAT is a prediction model that discovers the graph dependency relationships among multivariate and acts as a sample generator to confuse AE, while AE reconstructs the samples and acts as a discriminator that distinguishes a real sample from a generated one. A novel adversarial training between SGAT and AE is applied to amplify the errors of anomalies such that the prediction performance of SGAT is improved and the overfitting of AE is avoided. In addition, we aggregate the prediction error, the reconstruction error, and the adversarial error for anomaly detection, and develop a graph based anomaly interpretation method that locates the root causes from both local and global perspectives. Extensive experiments with five real-world data offer evidence that the proposed solution SGAT-AE is capable of achieving better performance when compared with the state-of-the-art proposals.