加载中...
本文提出 TIM(Transaction Intent Mining)——一个基于多主体(Meta-Level Planner、Perspective-Specific Domain Experts、Question Solvers、Cognition Evaluator)的 LLM 框架,用以联合多源链上/链下数据进行 DeFi 交易意图的多视角推断,构建了基于 Grounded Theory 的 DeFi 交易意图分类并通过实验验证 TIM 在召回、精确率与 F1 上均显著优于传统机器学习、单 LLM 与单代理基线,从而提升意图可解释性与可验证性 .
| Method | Recall | Precision | F1-micro |
|---|---|---|---|
| Naive Bayes | 0.76 | 0.36 | 0.49 |
| SVM | 0.59 | 0.50 | 0.54 |
| Decision Tree | 0.65 | 0.44 | 0.52 |
| XGBoost | 0.68 | 0.55 | 0.61 |
| CNN+sigmoid | 0.62 | 0.58 | 0.62 |
| Single LLM | 0.33 | 0.28 | 0.30 |
| Single Agent | 0.38 | 0.45 | 0.41 |
| w/o MP | 0.62 | 0.65 | 0.63 |
| w/o DE | 0.40 | 0.48 | 0.44 |
| w/o QS | 0.32 | 0.30 | 0.31 |
| w/o CE | 0.72 | 0.45 | 0.55 |
| TIM (default) | 0.78 | 0.72 | 0.75 |
| w/ grok-3-mini | 0.72 | 0.70 | 0.71 |
| w/ grok-3 | 0.80 | 0.69 | 0.74 |
| w/ gpt-4o-mini | 0.73 | 0.69 | 0.71 |
| w/ gpt-4o | 0.77 | 0.68 | 0.72 |
| w/ qwen3-30b | 0.69 | 0.67 | 0.66 |
| w/ LangGPT | 0.75 | 0.71 | 0.73 |
| Intent Code | Recall | Precision | F1-micro |
|---|---|---|---|
| A1 (Spot Trading Profit) | 1.00 | 0.73 | 0.84 |
| A2 (Leveraged Trading Profit) | 0.67 | 0.53 | 0.59 |
| A3 (Long-term Holding) | 0.82 | 0.73 | 0.77 |
| A4 (Arbitrage) | 0.60 | 0.56 | 0.58 |
| A5 (Provide/Create Liquidity Pool) | 0.53 | 0.55 | 0.54 |
| A6 (Participating in Lending) | 0.85 | 0.60 | 0.70 |
| A7 (Yield Aggregation) | 0.72 | 0.68 | 0.70 |
| A8 (ETH Liquid Staking) | 0.57 | 0.52 | 0.54 |
| A9 (DeFi Governance Token Staking) | 0.88 | 0.90 | 0.89 |
| A10 (Compound Liquid Staking) | 0.56 | 0.72 | 0.29 |
| A11 (Participating in Airdrops) | 0.30 | 0.28 | 0.29 |
| A12 (Participating in Presales) | 0.55 | 0.60 | 0.57 |
| A13 (Using Secure Wallets) | 0.92 | 0.80 | 0.86 |
| A14 (Permission Management) | 1.00 | 0.52 | 0.68 |
| A15 (Purchasing Insurance) | 0.82 | 0.74 | 0.78 |
| A16 (Stop-loss Strategies) | 0.53 | 0.43 | 0.47 |
| A17 (Hedging Strategies) | 0.42 | 0.45 | 0.43 |
| A18 (Voting) | 1.00 | 1.00 | 1.00 |
| A19 (Proposals) | 0.67 | 0.33 | 0.44 |
| A20 (Delegating Voting Rights) | 1.00 | 1.00 | 1.00 |
| A21 (Vulnerability Reporting) | 0.50 | 0.50 | 0.50 |
元数据与概览(引言与报告概览)
逐节深度解读
图表与图像深度解读
估值分析
风险因素评估(来自报告识别)
批判性视角与细微差别(审慎观点)
结论性综合
参考(指向报告中关键页以便溯源)
(注:以上所有对原文事实、数据、表格或图像的引用均基于原始文档中明确陈述的页码,相关句末已标注页码以便溯源。)
2511.15456.pdf