Know Your Intent: An Autonomous Multi-Perspective LLM Agent Framework for DeFi User Transaction Intent Mining

Battle Count: 99

The paper is primarily about understanding user transaction intents in DeFi rather than direct trading strategy development. However, understanding user intent (spot trading, arbitrage, leveraged trading, liquidity provision, staking) provides valuable behavioral signals that could inform quantitative trading strategies, market microstructure analysis, and risk assessment. The intent taxonomy covers trading strategies (A1-A4), yield farming (A5-A7), and risk management (A16-A17) which are directly relevant to DeFi quantitative trading. The framework could serve as a data enrichment layer for trading systems.

High complexity due to multi-agent architecture with 4 distinct agent types (Meta-Level Planner, Domain Experts, Question Solvers, Cognitive Evaluator), parallel execution, reflective multimodal data retrieval from blockchain nodes and web sources, structured task decomposition, memory management, and coordination mechanisms. Requires integration with blockchain APIs, web scraping, LLM APIs, and careful prompt engineering. The system involves multiple LLM calls per transaction analysis, JSON formatting of blockchain data, and context optimization strategies.

Code and prompts are available on GitHub (https://github.com/c0mm4nd/TIM). The paper provides detailed descriptions of the agent architecture, prompts (CRIPSE-style), experimental setup (8x RTX 3090 GPUs, Arch Linux), and transaction sources. However, the full annotated dataset of 600 transactions is not explicitly shared, and the system relies on external APIs (xAI Grok, OpenAI, OpenRouter) which may change over time. The intent taxonomy is referenced from prior work [16].