TLDR: HCLA is a human-centered multi-agent AI system (Parsing, Detection, Explanation) that uses LLMs and XGBoost to detect anomalous digital asset transactions. It allows non-experts to query in natural language, provides clear narrative explanations for anomalies, and enhances accessibility, interpretability, and trust compared to traditional opaque systems. It was validated on a Bitcoin-mixing dataset and showed improved user comprehension and confidence, despite some limitations in latency and domain adaptation.
In the complex and rapidly evolving world of decentralized finance, monitoring for suspicious activities like anomalous digital asset transactions presents significant challenges. Traditional detection systems often operate as “black boxes,” requiring specialized expertise to understand and use. This opacity can hinder trust and make it difficult for non-experts to verify findings.
Addressing this critical need, researchers have introduced HCLA, a Human-Centered LLM-Agent system designed to make anomaly detection in digital asset transactions more accessible, interpretable, and trustworthy. HCLA stands for Human-Centered LLM-Agent, and it aims to bridge the gap between sophisticated AI models and human understanding. You can read the full research paper here.
What is HCLA?
HCLA is a multi-agent system that orchestrates three distinct AI agents in a conversational workflow: a Parsing Agent, a Detection Agent, and an Explanation Agent. This modular design allows non-experts to interact with the system using natural language, inspect structured analytics, and receive clear, context-aware explanations for detected anomalies.
The core idea behind HCLA is to empower individuals to steer anomaly analysis without needing to master complex technical pipelines. Each step of the process is designed to be auditable, enhancing transparency and building trust in financial forensics.
How HCLA Works: The Three Agents
The system operates through a continuous interaction loop:
- Parsing Agent (ChatGPT): This agent takes everyday, free-form user requests (e.g., “Analyze transactions from my wallet over the past week”) and converts them into a structured JSON schema. This schema contains machine-readable variables like time windows, wallet IDs, and value ranges, making the user’s intent understandable to the detection model.
- Detection Agent (XGBoost): Once the input is structured, this agent computes anomaly probabilities for transactions. It uses temporal, transactional, and graph-connectivity features to identify suspicious patterns. While XGBoost is used as the baseline in the prototype, the system’s modularity allows for easy integration of other detection models like Graph Neural Networks (GNNs) or temporal models.
- Explanation Agent (Gemini): This agent is crucial for interpretability. It translates the numerical anomaly scores and underlying features from the Detection Agent into plain-language reasoning. For instance, it might explain that a transaction has a high anomaly score “due to repeated high-value transfers to unverified counterparties during off-peak hours.” Users can then ask follow-up questions, and the system continues the conversation, refining its analysis.
These agents communicate using JSON messages under a central LLM controller, ensuring that all parts of the system remain semantically aligned. This architecture promotes transparency, traceability, and adaptability, which are key characteristics of human-centered AI.
Key Advantages and Contributions
HCLA offers several significant benefits:
- Accessibility: Non-experts can issue queries in natural language without needing to write code or understand complex data filters. The system autonomously handles the conversion of natural language into structured data.
- Interpretability: Instead of just providing numerical scores, HCLA offers narrative explanations that link anomaly scores to concrete patterns, such as transaction frequency, counterparty details, or timing. This helps users understand not just what is anomalous, but why.
- Trust and Auditability: The multi-agent structure makes each step of the reasoning process visible and queryable. Users can trace the logic from parsing to detection to explanation, allowing them to verify consistency and replay the reasoning. This transparency is vital for building trust, especially in regulated financial environments.
Performance and User Study
The system was evaluated using a labeled Bitcoin-mixing dataset from Wasabi Wallet (2020–2024), comprising over 380,000 transactions. The baseline XGBoost detector achieved strong performance metrics, including an accuracy of 0.9159, precision of 0.9317, recall of 0.9159, and an F1-score of 0.9209, with an average response latency of less than 2 minutes per query.
A simulated user study with a panel of AI and Digital Assets experts compared HCLA’s narrative explanations against traditional numerical dashboards. Participants consistently rated HCLA’s explanations significantly higher in both trust and clarity. This empirical validation underscores the effectiveness of the human-centered approach in improving user understanding and confidence.
Identified Limitations and Future Directions
Despite its successes, the research identified several limitations:
- Computational Cost and Latency: The use of Large Language Models (LLMs) introduces computational overhead, leading to some response delay, which might limit real-time monitoring in high-frequency transaction streams.
- Domain-Specific Adaptation: General-purpose LLMs like ChatGPT and Gemini occasionally produced ambiguous interpretations or inconsistent terminology. Fine-tuning these models with blockchain and finance-specific data is expected to improve accuracy.
- Scalability Constraints: While effective on batch-processed datasets, scaling HCLA to continuous blockchain streams will require asynchronous orchestration and caching mechanisms.
Future research aims to address these limitations by incorporating real-time monitoring, integrating external multimodal data sources (like logs and news streams), and conducting large-scale user validation studies to further generalize findings on trust and interpretability.
Also Read:
- Decentralized Auditing Framework Enhances Trust in Large Language Model Reasoning
- LAFA: Enabling Natural Language Queries for Private Data Analytics with LLM Agents
Conclusion
The HCLA framework represents a significant step forward in making anomaly detection in digital asset ecosystems more human-centered. By integrating LLM-based parsing and explanation layers with robust detection models, it transforms a complex technical task into an interactive, understandable, and trustworthy process for non-expert users. This approach enhances transparency and adaptability, paving the way for more accountable AI systems in digital finance.
