%Aigaion2 BibTeX export van HES SO Valais Publications
%Sunday 03 May 2026 08:42:23 AM

@INPROCEEDINGS{,
     author = {Trerotola, Mario and Calvaresi, Davide},
      month = oct,
      title = {AI-Driven Multi-Agent Systems for Automated Regulatory Analysis of Crypto Projects},
  booktitle = {IEEE International Conference on Metrology for eXtended Reality, Artificial Intelligence and Neural Engineering (IEEE MetroXRAINE)},
       year = {2025},
  publisher = {IEEE},
   location = {ANCONA},
   abstract = {The rapid expansion of the crypto-asset market is stretching supervisory authorities and institutional investors, and although MiCAR mandates comprehensive disclosures, the sheer volume and heterogeneity of project whitepapers push regulators toward a largely passive, alert-driven form of oversight that
leaves little room for timely, preventive intervention. To address this challenge, this article presents an AI-driven multi-agent system (MAS) that automates Markets in Crypto-Assets Regulation (MiCAR) due diligence by synergizing large language models (LLMs) with specialized software agents. This MAS features a
supervisory agent orchestrating specialized peers for evidence retrieval, document normalization, semantic extraction, and rule-based verification, culminating in an auditable compliance report and an associated knowledge graph. Key techniques include retrieval-augmented generation (RAG) for contextualizing legal
texts, self-critique prompting to enhance LLM reliability, and containerized microservices for scalable deployment. Evaluated on a diverse corpus of public whitepapers, the system significantly reduces processing latency and expert workload, achieving compliance assessment accuracy comparable to human experts for critical MiCAR requirements. Furthermore, the system supports longitudinal monitoring by dynamically incorporating regulatory updates into its rule repository, ensuring ongoing alignment with evolving MiCAR standards. The MAS architecture’s transparent division of labor enables fault isolation, parallel processing,
and human-in-the-loop validation, providing superior robustness and interpretability over conventional monolithic AI models. By translating complex legal obligations into reproducible computational workflows, our approach advances Regulatory Technology (RegTech), offers actionable intelligence to market participants, and fosters a more transparent and accountable digital finance ecosystem. Future work will focus on multilingual capabilities and enhancing adversarial robustness against deceptive disclosures.}
}