This paper addresses the challenges of evidence dispersion, attention shift, and inference inconsistency in long text and multi-document reasoning scenarios. It investigates the problems that large language models often encounter under conditions of extremely long contexts, multi-source information conflict, and structural complexity, and proposes a hierarchical, curriculum-based fine-tuning algorithm framework. This framework organizes the input into a hierarchical structure of questions and multi-document contexts. At the representation level, it constructs a three-level convergence path of tokens, fragments, and documents to form structured memory and mitigate semantic drift caused by context expansion. At the reasoning level, it introduces a question-guided evidence scoring and weight aggregation mechanism to achieve differentiable selection across document fragments and global evidence vector construction, thereby strengthening the alignment of key evidence and suppressing redundant interference. At the training organization level, it employs a curriculum strategy, progressively scheduling samples according to difficulty levels, enabling the model's capabilities to gradually transition from local consistency to cross-document evidence integration and overall induction. Comparative experimental results show that this method exhibits more stable performance in multi-document evidence localization and inference consistency evaluation, validating the effectiveness of hierarchical modeling and curriculum scheduling in shaping multi-document reasoning capabilities.

Identifying abnormal behaviors plays a vital role in ensuring cloud infrastructure remains secure and operationally stable. Conventional methods that aggregate data at a single location create substantial privacy concerns, especially within shared cloud platforms hosting multiple organizations with confidential operational information. This paper proposes HierFedDP, a hierarchical federated learning framework integrated with a two-stage differential privacy mechanism for privacy-preserving anomaly detection in cloud services. Our approach employs a three-tier architecture consisting of local clients, edge servers, and a central cloud server, where clients apply local differential privacy to their updates before transmission. We introduce an edge aggregation frequency parameter that enables edge servers to perform multiple local aggregation rounds before communicating with the central cloud. Experiments on the CICIDS2017 dataset demonstrate that HierFedDP achieves detection performance comparable to standard local differential privacy approaches while reducing wide-area network (WAN) communication overhead by 49%. This significant communication reduction, achieved without sacrificing privacy guarantees or detection accuracy, makes HierFedDP particularly suitable for bandwidth-constrained, geo-distributed cloud deployments.