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

🕒 Published (v1): 2025-11-19 14:15 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

TIM (Transaction Intent Mining) is a multi-layered LLM agent framework for inferring DeFi user transaction intents from opaque on-chain hex logs and off-chain context. A Meta-Level Planner dynamically spawns perspective-specific Domain Expert agents, each of which decomposes its task into sub-questions solved by atomic Question Solver agents, with a Cognitive Evaluator performing post-hoc hallucination filtering. TIM achieves F1-micro 0.75 versus 0.41 for the best single-agent baseline.

Problem¶

DeFi transactions are recorded as hexadecimal logs and low-level smart contract calls that carry no human-readable semantics. Traditional ML approaches (Naive Bayes, SVM, XGBoost) rely on shallow structural features and miss complex multi-step intent patterns. Single LLMs and ReAct-style single-agent frameworks suffer hallucination and context-dependency failures when processing structured on-chain data combined with off-chain market signals. No existing method provides an unsupervised multi-label intent assignment system grounded in a principled DeFi intent taxonomy.

Method¶

TIM is formalized as \(M = \{A_1, A_2, \ldots, A_n; \mathcal{C}\}\) where each \(A_i\) is an LLM-based agent and \(\mathcal{C}\) is the coordination mechanism. The pipeline has four layers:

Meta-Level Planner (MP): Given a transaction and a pre-built intent taxonomy (\(I\) = 21 categories across 3 dimensions), MP dynamically generates a set of analysis perspectives (e.g., smart contract analysis, temporal context, market dynamics, protocol background) as JSON-structured expert configurations with prompts and tool suggestions.
Perspective-Specific Domain Experts (DE): One DE per perspective runs in parallel; each applies a divide-and-conquer strategy, producing a structured plan (objectives, to-do items, prompts) and spawning one Question Solver per sub-question.
Question Solvers (QS): Atomic single-responsibility agents operating under the ReAct paradigm—fetch data, reflect, re-fetch or think, summarize. QS agents use live multimodal data sources (blockchain node raw data, on-chain analytics APIs, web search, page browsing). Results accumulate in a sequential buffer passed to the next QS.
Cognitive Evaluator (CE): Aggregates all DE reports and filters intents along two axes—(i) verifiability of evidence and (ii) relevance to user intent—producing ranked final intent labels \(I_{tx}\).

The intent taxonomy (21 labels, 3 core categories: Investment/Speculation, Personal Risk Control, Project Governance) was derived via Grounded Theory from semi-structured interviews and 51 DeFi protocol documents, reaching theoretical saturation at 110 open codes → 21 axial categories → 9 themes.

Key Contributions¶

End-to-end autonomous DeFi transaction intent mining workflow with no manual intervention at inference time.
A verified DeFi user intent taxonomy (21 labels) built via formal Grounded Theory methodology.
Meta-Level Planner design for dynamic, transaction-adaptive perspective generation rather than static pipeline stages.
Reflective multimodal live data retrieval with iterative self-critique and source-switching.
Cognitive Evaluator as a structured post-hoc hallucination suppression layer with dual-criterion scoring.

Results¶

All metrics are F1-micro / Recall / Precision on an expert-labeled Ethereum DeFi transaction dataset:

TIM (Grok-2): F1 0.75, Recall 0.78, Precision 0.72
Single Agent (ReAct + tools): F1 0.41 — TIM is +83% relative F1
Single LLM (Grok-2, prompt only): F1 0.30
Best ML baseline (XGBoost): F1 0.61
Ablation — w/o MP: F1 0.63; w/o DE: 0.44; w/o QS: 0.31; w/o CE: 0.55
Performance is robust across model scales: Grok-3 F1 0.74, GPT-4o F1 0.72, Grok-3-Mini F1 0.71, Qwen3-30B F1 0.66
Per-label highlights: A18 (Voting) and A20 (Delegating Voting) achieve F1 1.00; A11 (Airdrops) F1 0.29, A10 (Compound Liquidity Mining) F1 0.29, A16 (Stop-loss) F1 0.47

Limitations¶

Complex, multi-step, or environmentally-dependent intents (airdrops, compound liquidity, hedging, stop-loss) remain poorly identified—likely requires domain-fine-tuned LLMs.
Dataset scope is restricted to Ethereum mainnet, mainstream protocols (Uniswap, Aave, Compound) and tokens, through March 2025; generalization to other chains or niche protocols is untested.
The intent taxonomy was grounded in interview data from a bounded set of 51 protocols; saturation does not guarantee completeness as the DeFi ecosystem evolves.
Latency and cost of spawning multiple parallel DE+QS agent chains at inference time are not evaluated; practical throughput at scale is unknown.
LangChain/AutoGen were rejected in preliminary experiments but no quantitative comparison to these framework baselines is provided.

Relevance to Harnesses / Meta-Harnesses¶

TIM is a concrete instantiation of the meta-harness pattern: the Meta-Level Planner acts as a dynamic orchestrator that generates agent configurations at runtime rather than hard-coding a static pipeline, directly paralleling meta-harness designs where a top-level controller spawns and configures specialized sub-harnesses based on input context. The three-layer decomposition (planner → domain expert → atomic solver) mirrors the hierarchical task decomposition seen in general-purpose harness architectures, and the Cognitive Evaluator functions as a structured quality-gate layer—a pattern increasingly common in production harnesses to reduce hallucination without requiring human-in-the-loop verification. The framework's demonstration that dynamic perspective generation by a planner agent substantially outperforms both static single-agent and fixed-pipeline designs (+83% F1 over ReAct single-agent) provides empirical evidence for the value of the meta-harness abstraction in domain-specific high-noise tasks.