Decoding User Intent in DeFi’s Complex World

Author: Denis Avetisyan

A new framework uses the power of artificial intelligence to accurately understand the reasoning behind complex transactions on decentralized finance platforms.

The Transaction Intent Mining (TIM) framework establishes a process where transaction parameters and an intent taxonomy guide domain experts in a parallel, multi-perspective analysis-utilizing live data and question-solving sequences-to generate and cognitively validate reports, ultimately delivering a final report detailing substantiated transaction intents and their supporting explanations.

This paper introduces TIM, an autonomous multi-agent system leveraging large language models for improved DeFi user transaction intent mining through collaborative reasoning and multimodal data integration.

Decentralized Finance (DeFi) presents a paradox: while offering unprecedented transparency through blockchain data, understanding why users engage in complex transactions remains opaque. This challenge motivates ‘Know Your Intent: An Autonomous Multi-Perspective LLM Agent Framework for DeFi User Transaction Intent Mining’, which introduces TIM, a novel system that leverages a multi-agent approach with large language models to autonomously and accurately infer user motivations from on- and off-chain data. TIM significantly outperforms existing methods by integrating diverse perspectives and mitigating LLM hallucinations, offering a more reliable understanding of complex blockchain activity. Could this framework unlock new levels of insight into DeFi user behavior and ultimately foster a more secure and user-centric ecosystem?

Decoding the On-Chain Signal: The Imperative of Intent Discovery

The burgeoning landscape of Decentralized Finance (DeFi) is characterized by a constant stream of on-chain transactions, creating a data-rich environment unlike any seen in traditional finance. However, simply having this data isn’t enough; a core challenge lies in deciphering the underlying intent behind each interaction. These transactions, while publicly recorded, often lack explicit labels indicating the user’s goals – are they engaging in arbitrage, providing liquidity, or speculating on a price movement? Without understanding why a user is interacting with a protocol, it becomes incredibly difficult to accurately assess risk, perform meaningful market analysis, or ensure regulatory compliance. The sheer complexity of DeFi protocols, with their nested smart contracts and multi-step processes, further obscures these intentions, demanding innovative approaches to extract actionable insights from the raw transaction data.

Current analytical techniques often fall short when applied to the intricacies of Decentralized Finance (DeFi). These methods, designed for traditional financial systems, struggle with the composability and rapid innovation characteristic of DeFi, leading to misinterpretations of on-chain activity. Consequently, accurate risk assessment becomes difficult, as the interconnectedness of smart contracts and liquidity pools obscures potential vulnerabilities. Market analysis suffers from an inability to discern genuine user behavior from automated trading or flash loan manipulations. Furthermore, regulatory compliance is hampered by the lack of tools capable of effectively tracing transactions and identifying illicit activities within this complex ecosystem, creating a significant need for novel approaches to understanding DeFi interactions.

Inferring a user's profit-driven spot trading intent necessitates a complex analysis combining on-chain transaction data, semantic summarization, and external market pricing. — Inferring a user’s profit-driven spot trading intent necessitates a complex analysis combining on-chain transaction data, semantic summarization, and external market pricing.

A Multi-Agent System for Unveiling Transactional Intent

The Transaction Intent Mining (TIM) framework utilizes a Multi-Agent System (MAS) to analyze Decentralized Finance (DeFi) transactions. This approach decomposes the complex task of intent discovery into smaller, manageable sub-tasks handled by independent agents. Each agent operates autonomously, communicating and collaborating to achieve a holistic understanding of the transaction’s purpose. The MAS architecture allows for modularity and scalability, enabling the framework to adapt to the evolving landscape of DeFi protocols and transaction types. By distributing analytical responsibilities, TIM facilitates a more efficient and accurate interpretation of on-chain activity compared to monolithic analysis approaches.

The Transaction Intent Mining (TIM) framework utilizes a multi-agent system comprised of three distinct agent types to analyze DeFi transactions. The Meta-Level Planner coordinates the overall analysis process, determining the necessary steps and allocating tasks to other agents. Perspective-Specific Domain Experts provide specialized knowledge regarding specific DeFi protocols, token contracts, or transaction types, enabling focused interpretation. Finally, Question Solvers address specific inquiries about a transaction – such as identifying the purpose of a swap or the risk associated with a liquidity pool – by leveraging the insights provided by the Domain Experts and the orchestration of the Planner. This division of labor allows TIM to decompose complex transactions into manageable components and extract nuanced intent information.

The agents within the Transaction Intent Mining (TIM) framework utilize Large Language Models (LLMs) to process and interpret the semantic content of DeFi transactions. These LLMs facilitate natural language processing of transaction-related data, including smart contract interactions and associated metadata. This capability extends beyond simple keyword recognition to encompass contextual understanding, allowing the agents to discern the underlying intent of transactions based on the sequence of operations and the specific parameters involved. Specifically, LLMs enable the agents to translate complex on-chain activity into human-readable descriptions and infer the user’s goals, even when those goals are not explicitly stated within the transaction itself.

The intent mining sequence diagram details the actions performed during the case study.

Establishing a Common Language: The DeFi Intent Taxonomy

A standardized DeFi Transaction Intent Taxonomy is essential for reliable data analysis due to the complexity and ambiguity inherent in on-chain activity. Without a consistent framework, interpreting user actions – such as swapping tokens, providing liquidity, or borrowing assets – is subject to individual bias and inconsistent labeling. This inconsistency hinders accurate aggregation of data, limiting the effectiveness of analytical processes like identifying trends, detecting anomalies, or profiling user behavior. A comprehensive taxonomy provides predefined categories and criteria for classifying transactions based on their intended purpose, enabling researchers and developers to consistently interpret and compare data across different platforms and time periods, ultimately improving the accuracy and scalability of DeFi analytics.

The DeFi Transaction Intent Taxonomy is constructed using Grounded Theory, a systematic methodology focused on deriving theory from analysis of qualitative data. This iterative process begins with open coding of transaction data – identifying initial concepts and patterns without predefined assumptions. These codes are then refined through axial coding, establishing relationships between concepts and grouping them into broader categories representing potential user intents. Selective coding further integrates these categories to form a core narrative, continuously tested and refined against new data. This cyclical approach ensures the taxonomy is empirically grounded in observed transaction behavior rather than driven by theoretical preconceptions, facilitating a robust and adaptable framework for intent classification.

The DeFi Intent Taxonomy refinement process utilizes multiple analytical perspectives to ensure comprehensive categorization of user actions. Smart Contract Analysis involves dissecting the code logic of interacted-with contracts to determine the precise function calls and intended outcomes. Temporal Context Analysis examines the sequence of transactions and time intervals between them to identify patterns indicative of specific strategies, such as arbitrage or liquidation. Finally, Market Dynamics Analysis incorporates external data, including price movements, liquidity pool sizes, and gas costs, to contextualize transactions and improve the accuracy of intent classification. The integration of these analytical methods allows for a multi-faceted evaluation, increasing the robustness and reliability of the taxonomy.

Ensuring Trustworthy Analysis: The Cognition Evaluator

To address the inherent risk of Large Language Model (LLM) hallucination – the generation of factually incorrect or nonsensical information – the TIM framework integrates a Cognition Evaluator (CE). This component functions as a critical safeguard against unreliable outputs by actively monitoring the analyses produced by other agents within the system. The CE is specifically designed to identify and flag instances where LLM-generated content deviates from established facts or logical reasoning, thereby minimizing the propagation of inaccurate insights and bolstering the overall dependability of the TIM framework’s results.

The Cognition Evaluator (CE) functions as a critical review stage within the TIM framework, systematically assessing output reports generated by other agents. This assessment centers on two key criteria: adherence to the pre-defined taxonomy and consistency with the broader contextual understanding established during the information-gathering process. The CE verifies that identified insights are appropriately categorized according to the taxonomy, and that reported relationships between data points are logically sound given the established context. This process involves evaluating the presence of unsupported claims, logical fallacies, and inconsistencies in the agent’s analysis, ultimately ensuring the reported findings are both structurally sound and contextually relevant.

The Cognition Evaluator (CE) utilizes Large Language Models (LLMs) to perform consistency checks on the analytical outputs of other agents within the TIM framework. This process involves evaluating statements against the established taxonomy and assessing internal coherence of the report, as well as external consistency with the provided contextual information. Identified inconsistencies, such as logical fallacies, contradictory statements, or taxonomic misclassifications, are flagged as potential inaccuracies. This flagging mechanism allows for human review or automated correction, ultimately enhancing the reliability and trustworthiness of the insights generated by the system.

Beyond Insight: Implications and the Future of DeFi Understanding

The Transaction Intent Mining (TIM) framework emerges as a significant advancement in deciphering the complex landscape of Decentralized Finance (DeFi). This tool moves beyond simple transaction tracking, offering a nuanced understanding of user intent through the analysis of on-chain activity. By identifying patterns in how users interact with smart contracts, TIM facilitates proactive risk management for DeFi platforms, enabling the early detection of potentially malicious activity and the mitigation of financial vulnerabilities. Furthermore, TIM’s capabilities extend to enhancing regulatory compliance, providing a transparent and auditable trail of user behavior that addresses growing concerns regarding illicit finance and market manipulation within the DeFi ecosystem. The ability to accurately categorize user intentions-whether for legitimate trading, yield farming, or exploitative practices-represents a crucial step toward fostering a more secure and trustworthy DeFi environment.

The Temporal Intent Model (TIM) distinguishes itself by synthesizing data from both on-chain blockchain activity and off-chain, real-world sources, creating a uniquely comprehensive understanding of decentralized finance (DeFi) interactions. This integration moves beyond simply tracking transactions; it allows TIM to correlate user behavior with external factors like news events, social media sentiment, and macroeconomic indicators. By bridging these traditionally disparate datasets, the model can discern not just what actions users are taking, but also why, providing insights into the motivations driving market trends and individual intent. This holistic view enhances the ability to anticipate shifts in user behavior, identify potentially malicious activity, and ultimately, build more robust and responsive DeFi systems.

Evaluations reveal that the Temporal Intent Model (TIM) consistently surpasses the performance of established baseline methods in discerning user intent within decentralized finance (DeFi). Across a spectrum of intent complexities, TIM achieves demonstrably higher precision, indicating fewer false positives in intent identification. Simultaneously, the model exhibits improved recall, signifying a greater ability to correctly identify all instances of a given intent. This combined strength is further validated by consistently superior F1-micro scores, a harmonic mean of precision and recall, which provides a balanced measure of the model’s overall effectiveness. These results confirm TIM’s capability to provide more accurate and comprehensive insights into user behavior than currently available techniques, offering substantial benefits for risk assessment and fraud prevention.

A significant barrier to widespread adoption of transaction intent monitoring (TIM) in decentralized finance (DeFi) has been the computational expense associated with analysis. Recent advancements have dramatically lowered these costs, achieving a reduction to under $0.1 per analysis – a substantial improvement from the initial approximate cost of $10. This optimization, achieved through algorithmic refinements and infrastructure improvements, unlocks the potential for real-time monitoring across a far broader range of transactions and users. The decreased financial burden not only facilitates more frequent analysis but also encourages the integration of TIM into existing DeFi platforms, fostering increased security, improved risk management, and more effective fraud detection across the ecosystem.

The presented framework, TIM, embodies a philosophy of systemic understanding. It doesn’t merely isolate individual transactions but orchestrates a collaborative reasoning process across multiple agents to discern user intent. This mirrors a core tenet of robust system design – that structure dictates behavior. As Ken Thompson observed, “Sometimes it’s better to keep it simple and not worry about all the edge cases.” TIM’s strength lies in its ability to navigate the complexities of DeFi by breaking them down into manageable components, much like a well-architected system. The multi-agent approach allows for a holistic view, recognizing that understanding intent requires considering the interconnectedness of on-chain data and smart contract interactions. This collaborative structure is key to outperforming existing intent mining methods.

Where Do We Go From Here?

The pursuit of deciphering user intent within decentralized finance, as exemplified by the TIM framework, reveals a fundamental truth: systems break along invisible boundaries – if one cannot clearly articulate the assumptions embedded within a transaction, pain is coming. Current methods treat on-chain data as a self-contained narrative, yet human action is rarely so neatly packaged. The next evolution necessitates a broader conception of context, integrating off-chain signals – social media, forum discussions, even the timing of transactions relative to world events – to build a more holistic understanding of user motivation.

However, simply accumulating more data is not the solution. The challenge lies in discerning signal from noise, and avoiding the creation of echo chambers within the multi-agent system. A critical area for future research is the development of mechanisms to actively challenge the prevailing interpretations of intent, fostering adversarial reasoning among the agents. Only through such internal friction can the system escape the trap of confirmation bias and approach a more objective understanding.

Ultimately, the true test of any intent mining system will not be its accuracy in classifying known transactions, but its ability to anticipate unforeseen actions. Structure dictates behavior; therefore, a deeper investigation into the underlying game-theoretic principles governing DeFi participation is essential. Only by modeling the rational (and irrational) forces driving user decisions can one hope to predict where the system will break – and, perhaps, build something more resilient in its place.

Original article: https://arxiv.org/pdf/2511.15456.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/