Introduction
The tokenization of Real-World Assets (RWAs) has emerged as a significant trend in decentralized finance, with tokenized U.S. Treasuries now representing a market exceeding $6 billion. Prominent examples include BlackRock’s BUIDL and Franklin Templeton’s BENJI. Despite their growth, these assets often operate within "walled gardens"—permissioned environments requiring Know Your Customer (KYC) and whitelisting. Consequently, the on-chain behavior of these institutional-grade assets has remained largely opaque to traditional blockchain analysis tools.
Existing methodologies often treat transaction networks as "flat" Euclidean graphs, failing to capture the inherent latent hierarchy of financial systems. In these systems, central institutional hubs (issuers and treasuries) coordinate flows to a vast periphery of retail users and automated bots. This paper introduces PoincaVec, a novel machine learning framework utilizing hyperbolic geometry to map these hierarchical structures and infer the economic roles of addresses with high precision.
Methodology
The authors propose a multi-stage pipeline for address role inference, moving beyond simple transfer analysis.
Contract Decoding and Functional Mapping
Raw smart contract calls are decoded into high-level functional "buckets" such as minting, redeeming, and bridging. This categorization allows the model to distinguish between routine value transfers and administrative or protocol-level interactions.
The PoincaVec Framework
The core innovation is the use of hyperbolic space (specifically the Poincaré ball) for address embeddings. Unlike Euclidean space, hyperbolic geometry expands exponentially as one moves away from the center, making it mathematically optimal for representing tree-like hierarchies. By measuring an address's "radius" in the Poincaré disk, the model derives a continuous proxy for its depth in the financial hierarchy: central institutional nodes sit near the origin, while peripheral retail actors reside near the edge.
Feature Engineering
To further refine role detection, the authors introduced the Liquidity-to-Average Ratio (LAR). This feature captures temporal liquidity behavior, aiding in the differentiation between passive long-term holders and high-frequency automated bots.
Empirical Dataset
The study analyzes a newly curated dataset of RWA transactions across multiple blockchains, including Ethereum, Arbitrum, Mantle, and Stellar. The dataset covers diverse market segments, from the institutional-heavy BUIDL to the retail-mixed USDY.
As shown in Table 1, the dataset encompasses a wide range of activity levels. For instance, Stellar-based BENJI shows massive transaction volumes driven by automated dividend reinvestments, while Ethereum-based BUIDL exhibits fewer, higher-value transactions.
Results and Discussion
The PoincaVec framework demonstrates superior performance in role inference compared to established baselines like Node2Vec and Role2Vec. On the RWA dataset, the model achieved an F1 score of 0.726 and an accuracy of 0.748.
Visualizing Behavioral Niches
The analysis reveals a stark bifurcation between institutional and retail activity. Institutional actors dominate high-value issuance and redemption on Ethereum, whereas retail users engage in mid-value swaps and transfers, primarily on Layer-2 networks like Mantle.
Figure 1 highlights the behavioral niche of BUIDL. The clustering in the high-value/low-frequency quadrant confirms its primary role as an institutional settlement asset. In contrast, tokens like USDY show broader dispersion, reflecting a more diverse user base.
Generalization and Anomaly Detection
The model's "curvature-aware" approach proved highly effective at identifying peripheral anomalies. When tested on external datasets like the AscendEXHacker dataset, PoincaVec achieved an F1 score of 0.840, significantly outperforming traditional Euclidean methods in spotting hackers and scammers who deviate from standard community patterns.
Conclusion
This research provides the first systematic, transaction-level analysis of the tokenized U.S. Treasury market. By leveraging hyperbolic geometry, the PoincaVec framework successfully decodes the hierarchical nature of RWA networks. Future work will focus on adapting the model to real-time anomaly detection as these assets become increasingly integrated into the broader DeFi ecosystem as collateral.
Summary updated