Intent, Actually: Why DeFi Needs a Mind‑Reader

A wallet is easy to watch and hard to understand.

That is the small comedy at the centre of DeFi analytics. Every transaction is public, every contract call can be inspected, every log can be dragged into a dashboard, and yet the actual question remains stubbornly human: what was this user trying to do? A swap may be a trade, a hedge, a liquidation defence, an arbitrage leg, a farming manoeuvre, or just someone clicking through a protocol interface with dangerous confidence. The chain shows the footprint. It does not provide the diary.

The paper behind this article, “Know Your Intent: An Autonomous Multi-Perspective LLM Agent Framework for DeFi User Transaction Intent Mining,” proposes TIM, a Transaction Intent Mining framework that uses a structured multi-agent LLM workflow to infer DeFi user transaction intents from on-chain and off-chain evidence.¹ The important word is infer. TIM is not a psychic engine for discovering a trader’s inner soul, mercifully. It is a system for producing evidence-supported labels over a predefined DeFi intent taxonomy.

That distinction matters. Without it, “intent mining” sounds like surveillance theatre with better branding. With it, the work becomes more interesting: a blueprint for converting messy blockchain behaviour into auditable semantic explanations.

The hard part is not classification; it is reconstructing context

Most transaction analytics begins with the visible object: a transaction hash, a sender address, a contract address, a value field, logs, traces, gas, and function signatures. That is useful, but it is not yet meaning.

A DeFi transaction may hide its purpose across several layers:

• the direct smart contract method;
• the sequence of previous and later transactions;
• the protocol’s economic design;
• market conditions around the transaction;
• token flows triggered indirectly through events;
• approvals, revocations, wrapper contracts, aggregators, and other little gifts from the user-experience department of the abyss.

The paper’s opening example is simple in intent but not simple in evidence: spot trading for profit may require analysing three related transactions, extracting their semantics, and incorporating external market price data. The obvious action is not enough. A transaction can look like a swap while functioning as one leg of a strategy.

This is why TIM is better understood as an analytical workflow than as a classifier. The contribution is not merely that an LLM is asked to assign labels. We have all seen that movie; the ending usually involves confident nonsense in bullet points. The paper’s main claim is architectural: DeFi intent analysis improves when the model is forced to decompose the investigation, retrieve relevant evidence, reason from multiple perspectives, and then pass its conclusions through an evaluator that filters for verifiability and intent relevance.

In other words, TIM tries to make the LLM behave less like a pundit and more like a junior analyst who has been given a checklist, tools, supervision, and no permission to vibe.

TIM turns one opaque transaction into several narrower investigations

The mechanism begins with a taxonomy. The paper uses a DeFi intent framework derived through grounded theory in earlier work, summarised into 21 intent labels under the broader idea of “risk-reward trade-off orientation.” The labels span investment or speculative profit-seeking, personal risk control and management, and project participation or ecosystem governance.

That taxonomy is not decoration. It defines what the system is allowed to infer. TIM is not generating arbitrary psychological explanations. It is assigning possible labels from a structured set: spot trading profit, leveraged trading profit, long-term holding, arbitrage, liquidity provision, lending, yield aggregation, ETH liquid staking, governance token staking, compound liquid staking, airdrop participation, presales, secure wallets, permission management, insurance, stop-loss, hedging, voting, proposals, delegated voting, and vulnerability reporting.

Once the label universe exists, TIM runs a deterministic multi-agent workflow. Its moving parts are easier to understand as a chain of narrowing responsibilities.

This structure is the paper’s real argument. A single LLM prompt sees a transaction and tries to answer the whole question at once. TIM instead asks: which parts of the world must be inspected before any answer deserves to exist?

The Meta-Level Planner may create perspectives such as smart contract analysis, temporal context analysis, market dynamics analysis, protocol background analysis, or even hacker behaviour analysis, depending on the transaction. Domain Experts then turn those perspectives into question chains. A smart-contract expert might ask which contracts were touched, which methods were called, and what effects those methods created. A temporal expert might inspect prior and later wallet activity. A market expert might look for price movements or context that makes a hedge or stop-loss plausible.

The Question Solvers perform the granular work: retrieve data, inspect it, decide whether more retrieval is needed, and summarise. The paper describes this as reflective multimodal live data retrieval: blockchain data is simplified into more machine-readable JSON-like text, HTML is reduced to main body text, and agent communications are kept in textual form. This is not glamorous, but it is exactly where many “AI for blockchain” demos quietly go to die. Hex logs do not become business insight just because an LLM was nearby.

Finally, the Cognition Evaluator reviews the outputs across two dimensions: whether factual claims are traceable to evidence, and whether those claims actually support an intent label. That second filter is important. A report can be factual and still irrelevant. DeFi analytics has no shortage of technically correct clutter.

The paper’s main evidence is a comparison, not a miracle

The experiment evaluates TIM on 600 Ethereum DeFi transactions, sampled across major protocols and tokens and labelled by domain experts. The dataset covers transactions up to March 2025, with sources including protocols and scenes such as Uniswap, Aave, Compound, SushiSwap, Lido, MakerDAO, Nexus Mutual, Pendle, Yearn, Gnosis Safe, ENS, and others.

The annotation process is worth pausing on because it sharpens the interpretation. Appendix D says the labels are inferred by experts, not directly reported by users. Ten annotators came from CEX AML, fund security, trading, development, and scholarly backgrounds. Each transaction was annotated by two independent annotators, with Cohen’s kappa of 0.7 before adjudication, and final labels were produced through expert consensus.

That means the benchmark measures agreement with external expert inference. It does not prove access to a user’s private motive. This is not a weakness so much as the only realistic basis for such a task. Blockchains do not ship confessionals with calldata.

The headline result: TIM reaches recall 0.78, precision 0.72, and F1-micro 0.75. It outperforms traditional machine-learning baselines, a single LLM, a single agent, and several ablated TIM variants.

The most revealing comparison is not TIM versus Naive Bayes. It is TIM versus the single LLM and single agent setups. A single LLM with basic transaction information performs poorly. A single tool-using agent improves but still trails traditional ML. TIM’s improvement comes from structured decomposition, not from merely attaching tools to a chatbot and calling it a platform.

That is a useful business lesson because many AI deployments still treat “agentic” as a procurement adjective. TIM suggests a harsher standard: agency becomes valuable when the workflow constrains responsibility, routes evidence, and evaluates claims.

The ablations show which parts of the machine are doing work

The paper’s ablation study is not a second thesis. Its purpose is to test whether TIM’s components are ornamental or necessary.

The most damaging removals are Question Solvers and Domain Experts. That is intuitive. If the system cannot break the task into questions and cannot assign granular evidence work, it falls back toward the same “one big answer” problem that makes LLM-based analysis fragile.

Removing the Cognition Evaluator is also revealing. The model without CE has recall 0.72 but precision 0.45, producing F1 0.55. That pattern is familiar: the system finds many possible intents but lets too much through. In business terms, it becomes noisy. And noisy intent analytics is not intelligence; it is just a dashboard with confidence issues.

The sensitivity analysis adds a smaller but useful point. TIM separates “creative” roles, such as planning and decomposition, from “executive” roles, such as solving and evaluation. The best reported temperature configuration uses moderate randomness for creative roles and deterministic output for execution roles. That is a good design instinct. Exploration belongs in hypothesis generation. Verification should be boring. The machines may be artificial, but the audit trail should not be jazz.

Per-label results reveal the actual boundary of intent inference

TIM performs best when the intent has a clear on-chain footprint. Voting and delegated voting both reach F1 1.00 in the paper’s per-label table. DeFi governance token staking reaches F1 0.89. Using secure wallets reaches F1 0.86. Spot trading profit reaches F1 0.84, though with lower precision than recall.

The weaker categories tell the more important story. Participating in airdrops reaches F1 0.29. Compound liquid staking is reported with weak F1. Stop-loss strategies reach 0.47. Hedging strategies reach 0.43. Proposals and vulnerability reporting are also difficult.

That distribution is exactly what we should expect. Some intents announce themselves in the method name or protocol action. Voting looks like voting because, at some level, it is voting. Delegation often leaves a crisp trace. Staking is usually recognisable through protocol-specific actions.

Airdrop farming, hedging, and stop-loss behaviour are different animals. They depend on sequences, market context, alternative possibilities, and sometimes behavioural interpretation that remains ambiguous even to humans. A wallet that sells after a price decline may be stopping loss, rebalancing, exiting a position, funding another wallet, responding to news, or simply panicking in the traditional human manner.

The paper’s value is therefore not that it solves ambiguity. It measures where structured evidence helps and where inference remains brittle.

That boundary is commercially important. A DeFi product team can use high-confidence intent categories for segmentation and UX analytics. A risk team can use lower-confidence categories as triage signals. A compliance team should not treat ambiguous labels as enforcement-grade facts. Different decisions require different evidentiary thresholds. This is where many analytics products, bless them, discover that a colourful label is not the same as a defensible conclusion.

The case study shows why mechanism beats benchmark worship

The paper’s case study illustrates TIM as a progressive reasoning chain. The Meta-Level Planner splits the task into smart contract analysis, temporal contextual analysis, and market dynamics analysis. Domain Experts generate questions. Question Solvers retrieve ABIs, traces, function signatures, and market information. The Cognition Evaluator then integrates and ranks likely intents.

The final case identifies a compound transaction intent involving governance token staking, liquidity provision, and spot trading. These are not guessed from a single surface-level call. Governance token staking is inferred through market and protocol knowledge. Liquidity provision is derived from yield logic implemented in the token’s smart contract code. Spot trading is determined by combining function signatures with subsequent sell actions.

This case study is exploratory. It does not prove general performance. Its purpose is interpretability: it shows how the system constructs a reasoned path from evidence to intent.

That matters because business adoption of agentic systems will not be won by leaderboard deltas alone. A risk analyst, product manager, auditor, or protocol governance team needs to know why an intent was assigned. “The model said so” is not an explanation; it is a liability wearing a hoodie.

TIM’s mechanism-first contribution is that it creates an inspectable intermediate structure: perspectives, questions, retrieved evidence, expert reports, and evaluator filtering. The workflow itself becomes part of the explanation.

The business value is semantic routing, not mind-reading

The obvious use case is wallet analytics: classify user behaviour more meaningfully than raw counts of swaps, approvals, and deposits. But the more interesting business value sits in routing decisions.

For DeFi protocols, intent labels could help identify where users struggle. If many wallets perform approval-management behaviour after interacting with a particular product path, the interface may be inducing fear or confusion. If liquidity provision is often followed by rapid withdrawal, incentives may be attracting mercenary capital rather than sticky participation. If governance delegation clusters around certain events, DAO tooling may need better explanation, coordination, or fraud detection.

For risk monitoring, intent-aware systems can prioritise review. A transaction that looks like ordinary liquidity movement may become more interesting if temporal and market context suggest hedging, liquidation defence, or exploit preparation. TIM does not directly prove that it can detect those scenarios in production, but it points to a richer alerting layer: not “this address interacted with a contract,” but “this sequence resembles a risk-management or speculative strategy under current conditions.”

For compliance and AML teams, the appeal is sharper but also more dangerous. Intent labels can improve triage by adding behavioural hypotheses to transaction monitoring. They should not be treated as ground truth. The paper’s labels are expert-inferred from external evidence; production labels would carry similar uncertainty. Used correctly, that is useful. Used badly, it becomes automated overclaiming with a subpoena problem.

For AI-agent platforms, TIM offers another lesson. As autonomous execution agents begin transacting on-chain, downstream systems will need to explain not only what agents did but why their actions were consistent with a declared strategy or risk policy. TIM’s architecture could become a model for post-trade explanation and governance: decompose the action, retrieve evidence, evaluate relevance, and produce a ranked account of intent.

The business pathway is therefore:

1. collect transaction data and related context;
2. retrieve live on-chain and off-chain evidence;
3. decompose the analysis into role-specific perspectives;
4. assign evidence-ranked intent labels;
5. route those labels into analytics, monitoring, UX design, or review workflows.

The paper directly supports steps 1 through 4 in an experimental setting. Step 5 is Cognaptus’ business inference. The bridge is plausible, but deployment would need calibration, cost control, latency targets, model governance, and clear thresholds for how each label may be used.

Deployment economics are part of the architecture, not an afterthought

The paper includes an unusually practical note on cost. Because agents retrieve and process large amounts of external data, context optimisation is essential. Without optimisation, inputs may exceed model context limits. The authors report that by formatting and removing machine-generated data such as hashes, the average cost of analysing one transaction intent using Grok-2 can be reduced from around $10 to less than $1. With a mini model, most transactions can reportedly fall below $0.10.

This is not a side detail. It is the difference between a lab prototype and a service someone might actually run.

The model-variant results also matter here. TIM with Grok-3 reaches F1 0.74, GPT-4o reaches 0.72, GPT-4o-mini reaches 0.71, Grok-3-mini reaches 0.71, Qwen3-30B reaches 0.66, and the default TIM result is 0.75 with Grok-2. The exact ranking is less important than the pattern: the framework appears less dependent on one frontier model than a prompt-only solution would be.

That suggests a useful implementation strategy. Spend expensive reasoning where ambiguity is high. Use cheaper models where the workflow is constrained, the taxonomy is narrow, and the evidence is clean. This is not glamorous either. It is also how serious AI systems become financially survivable.

Where TIM should not be overread

The paper is strongest as evidence for structured, evidence-grounded intent inference over a fixed DeFi taxonomy. It is weaker if read as a universal theory of user motivation.

There are four boundaries worth keeping visible.

First, the ground truth is expert annotation, not self-reported user intent. The paper is explicit about this in the appendix. Expert inference is a reasonable benchmark for transaction analytics, but it remains external interpretation.

Second, the dataset contains 600 Ethereum DeFi transactions. That is substantial enough for a focused study, but not enough to declare broad production reliability across every chain, protocol, market regime, and adversarial behaviour pattern. Cross-chain execution, account abstraction, private order flow, bridges, and intent-centric protocols could all complicate the evidence surface.

Third, ambiguous intents remain difficult. Airdrop participation, hedging, stop-loss, and compound liquidity strategies are precisely the categories where context is incomplete and labels overlap. TIM improves the workflow, but it does not abolish ambiguity. The universe continues to be rude.

Fourth, the Cognition Evaluator reduces hallucination risk but cannot eliminate it. It evaluates verifiability and relevance based on available reports and evidence. If retrieval misses important context, if protocol documentation is misleading, or if market interpretation is underdetermined, the evaluator may still approve an incomplete conclusion.

These boundaries do not weaken the paper’s practical relevance. They define it. TIM is best seen as a decision-support system for evidence-ranked behavioural interpretation, not as an oracle.

The real lesson: DeFi needs evidence architecture

The useful idea in TIM is not that DeFi needs bigger LLMs. It is that DeFi needs evidence architecture.

Raw transaction data is too low-level. Classical features are too shallow. Prompt-only LLMs are too unconstrained. Single agents with tools are too undisciplined. TIM’s result is interesting because it sits between these extremes: structured enough to be auditable, flexible enough to handle heterogeneous transactions, and evidence-aware enough to avoid some of the more theatrical hallucinations.

For businesses, the implication is clear. If you want AI to explain complex financial behaviour, do not begin by asking for a magical label. Begin by designing the investigation: what perspectives are required, what evidence counts, what claims must be verified, and which outputs are allowed to influence which decisions.

In DeFi, the chain tells us what happened. TIM’s contribution is a disciplined attempt to explain what it probably meant.

That is not mind-reading. It is something more useful: structured suspicion with receipts.

\ast\astCognaptus: Automate the Present, Incubate the Future.\ast\ast