Abstract: Money laundering poses a serious threat to global financial security and social stability. Anti-money laundering (AML) has become a strategic priority for maintaining financial integrity and imposes binding compliance obligations on financial institutions. As money laundering schemes become increasingly covert and sophisticated, conventional static rule-based monitoring is no longer sufficient, and intelligent techniques are becoming an important approach to AML modeling. Existing studies, however, remain scattered across machine learning, graph learning and large language models (LLMs), and a systematic review aligned with the structure of AML business tasks is still lacking. This survey systematically reviews research progress in intelligent AML techniques from four analysis granularities: entity, transaction, network and semantic. It first defines AML tasks hierarchically according to analysis granularity, specifying the analysis object, typical outputs and inter-level relationships at each level. It then reviews detection methods, including rule-based scoring, machine learning, deep sequential modeling and graph learning, for customer and account risk identification, transaction anomaly detection and fund-flow linkage analysis, analyzes the settings in which they apply, and identifies their limitations in exploiting unstructured text, adapting to label scarcity and ensuring interpretable decisions. To address these limitations, the survey further examines the application of LLMs to semantic-enhanced detection, knowledge augmentation and compliance reasoning, and to agent-driven investigation support and suspicious activity report generation; reliability issues including hallucination mitigation, grounding, auditability and traceability are also discussed. In addition, it compiles the public and synthetic data resources available in this field, summarizes evaluation metrics by task level, and identifies key challenges concerning data foundations, evaluation comparability, cross-granularity integration, interpretability and compliance, while outlining future directions in terms of evaluation standardization and data ecosystems, cross-granularity technical coordination, and auditable agents for human-AI collaborative investigation. Compared with existing surveys, this review organizes methods by analysis granularity rather than algorithm type so that technical solutions map directly onto AML business tasks, extends its scope from detection models to generative agents to cover the technical chain from risk identification to investigation support, and consolidates available datasets and evaluation resources, with a view to providing a reference for both academic research and engineering practice in intelligent AML.

摘要： 洗钱活动严重威胁全球金融安全与社会稳定，反洗钱（Anti-Money Laundering, AML）已成为各国维护金融秩序的战略性任务，并对金融机构合规运营提出刚性要求。随着洗钱手法日趋隐蔽化与智能化，传统基于规则的静态监测手段难以应对，智能技术正成为反洗钱建模的重要路径。但相关研究分散于机器学习、图学习与大语言模型等领域，尚缺乏与业务任务结构相对应的系统梳理。本文以主体、交易、网络与语义四类分析粒度为框架，系统综述反洗钱智能技术的研究进展。首先依据分析粒度对反洗钱任务作分层界定，明确各层级的分析对象、典型输出及其衔接关系。继而围绕客户与账户风险识别、交易行为异常检测与资金链路关联分析等任务，评述规则评分、机器学习、深度时序建模与图学习等检测方法，分析其适用条件，指出其在非结构化文本利用、标注稀缺适应性与决策可解释性上的不足。针对上述问题，进一步论述大语言模型在语义增强检测、知识增强与合规推理、智能体驱动的调查与报告生成等方面的应用进展，并讨论幻觉抑制与可追溯性等可靠性问题。此外，本文整理该领域的公开与合成数据资源，按任务层次归纳评价指标，并据此总结数据基础、评测可比性、多粒度融合与可解释合规等方面的关键挑战，展望评测标准化与数据生态、多粒度技术协同、可审计智能体与人机协同调查等方向。与既有综述相比，本文以分析粒度而非算法类型为主线组织方法，使技术方案与业务任务直接对应；将研究视野由检测模型延伸至生成式智能体，贯通从风险识别到调查支持的技术链条；并整理数据集与评测资源，以期为该领域的研究与工程实践提供参考。