Key takeaways

• Foundation models help defenders keep pace with adversaries who can automate, adapt, and move across chains quickly.
• Learning from unlabeled transaction histories helps alleviate blind spots; the model can learn general blockchain behavior before it is fine-tuned on known attribution. Furthermore, it also helps us measure where the labeled dataset differs from the broader on-chain population, so we can identify and mitigate potential selection bias.
• Fine-tuning on TRM's attribution inventory creates an additional attribution expansion path. High-confidence predictions on unlabeled addresses can be added to our attribution inventory if they meet rigorous evidence standards or turn into prioritized candidates for analyst review, depending on the use case.
• The model lets investigators search by behavior, not only by exact rule. Similarity and behavioral search can surface related addresses that hand-written queries may miss.
• Precision, calibration, and point-in-time evaluation matter because model outputs can shape investigative attention. Nothing from the model is put in front of a customer without rigorous validation and a communicated level of confidence.

‍

{{horizontal-line}}

‍

Criminal enterprises are already using AI and automation to scale fraud, iterate laundering patterns, and move across blockchains faster than manual review can follow. The next phase of blockchain intelligence is therefore not only about better analytics, but building defensive AI systems that can help investigators and compliance teams outpace adaptive adversaries.

Public blockchains give defenders a structural advantage: every transfer is timestamped, ordered, and visible. But the useful signal is rarely in one field or transaction. Timing, funding source, counterparties, concentration, graph position, and prior context can change what the same activity means. A simple exchange deposit may look routine on its own but different after a burst of small inbound transfers, funding from a high-risk venue, or inside a broader group of addresses with similar behavior.

This is where handcrafted analytics reach their limits. Metrics defined by hand — like transaction counts, volumes, counterparties, recency, concentration, and exposure — are valuable, but adversaries do not organize their behavior around them. The strongest signal often comes from how many weak signals interact.

That is why TRM is training its own transformer-based Blockchain Foundation Model. The model learns from labeled and unlabeled addresses and transaction histories: first learning broader behavioral patterns from public chain activity, then leveraging proprietary attribution to specialize that representation for intelligence workflows.

To our knowledge, TRM is among the first to apply transformer-based foundation models to blockchain intelligence, combining public chain transaction histories with one of the largest proprietary blockchain attribution datasets in the industry. The goal is to move beyond one-off rules and known-label lookups to using unlabeled activity to alleviate blind spots, learning what different addresses have in common from mixed signals, surfacing similar behavior through search and clustering, and using calibrated model confidence to support attribution expansion safely in analyst workflows — and only in customer-facing intelligence after rigorous validation.

The shift

‍

{{horizontal-line}}

Building a blockchain foundation model

The defining strength of transformer-based models is their ability to take in many complex, disparate signals at once and fuse them into a single, information-rich representation. In large language models (LLMs), that is what lets the model absorb a long stretch of text — the words, their order, and the surrounding context — and turn it into a representation expressive enough to power applications like coding.

A blockchain foundation model applies that same strength to on-chain activity. An address' behavior is made up of many individually weak, disparate signals: the timing of transfers, value moving in and out, the assets and transfer types involved, its counterparties, and where it sits in the broader transaction graph. The model learns to integrate all of these into an embedding of how an address behaves — so addresses that behave alike sit nearby in the vector space even when they share no simple hand-written rule.

The model leverages both public on-chain activity and TRM's labeled attribution inventory — one of the largest proprietary blockchain attribution datasets in the industry. This combination of public chain scale, proprietary attribution context, and transformer-based representation learning can be transformative for blockchain intelligence.

How the model works

The model starts from an address and its recent transaction history — for each event, whether value moved in or out, the asset and transfer type, the amount, the timing, and high-level context about the counterparty. Unlike a bank's customer-to-merchant record, on-chain activity is multi-party graph-shaped: addresses interact with exchanges, bridges, mixers, scam infrastructure, smart contracts, and each other. So alongside each address' own sequence of events, the model takes in enough structural context to know where that activity sits in the broader network.

The model then reads the on-chain activity and compresses it into an embedding. Those embeddings live in a shared vector space, which is the foundation layer. Every downstream application — category, illicit activity, entity prediction — and every discovery workflow, from similarity search to clustering, draws on the same representation rather than on a separate purpose-built model.

Learning from labeled and unlabeled data

The model learns in two stages. First, self-supervised pre-training. Similar to how an LLM learns by trying to predict the next word, the model learns blockchain behavior by predicting masked transactions — general behavior it can absorb without any labels. Then there’s supervised fine-tuning on attribution labels, where the model learns to label illicit activity, known entities, and broad categories.

From model output to analyst workflow

The model is most useful when its outputs map directly into the ways investigators already work: expanding attribution, finding related addresses, and searching for behavior that is hard to capture with a single rule. A high-confidence prediction can point to an unlabeled address worth reviewing. A search for similar behavior can reveal a related cluster. A broader behavioral search can surface activity that resembles a known typology, even when the exact counterparties, assets, or routes have changed.

Those workflows fall into three analyst-facing use cases:

1. Attribution expansion: The model flags unlabeled addresses that look similar to known categories, entities, or illicit activity patterns.
2. Similarity search: Analysts start with one address or cluster and find other addresses that behave the same way.
3. Behavioral search: Investigators search for patterns, not just exact rules: "addresses that look like exchange deposit infrastructure," "wallets with cash-out timing similar to this typology," "clusters near this entity but not yet attributed."

For example, in an internal review of a sanctioned cash-out network, TRM started from one known address and asked the model to find addresses that behaved like it instead of writing a rule for that exact path. The model surfaced other unlabeled addresses with the same cash-out profile: funds arriving from connected sources, then moving on to deposit infrastructure. One known case became a broader review set that was used by analysts to expand investigations — not because the addresses shared a counterparty, but because they behaved alike.

Model outputs are evaluated point-in-time: we can reconstruct what the model saw, what version produced a prediction, and what confidence it carried on the date it was made. An investigator who acted on a signal in March can show exactly what the system knew in March, not what it knows today.

Where this work goes next

The goal is not to build one model for one narrow task. It is to give investigators more ways to turn the transparency of the blockchain into more usable intelligence.

The same model can help analysts find addresses that behave like a known service, group related activity for review, identify unlabeled addresses that look similar to known entities or categories, and follow patterns as they move across chains. Over time, this creates a shared intelligence layer: one that learns from public chain activity, uses TRM's attribution context responsibly, and helps defenders understand emerging activity faster.

Defensive AI advantage

• See more activity: Public chain transparency
• Understand more context: Foundation-model representations
• Act earlier: Attribution expansion and discovery workflows
• Stay accountable: Calibration, review, and auditability

How TRM uses machine learning — and how we don't

For readers wondering how TRM uses machine learning: this work is the start of the journey, not a customer-facing model in production. The model helps investigators surface leads and similar behavior. Before any of that information would ever be put in front of a customer, TRM conducts several rigorous validation steps and communicates our level of confidence. Model output is a starting point for review, not a label we ship.

Public blockchains already contain a rich record of observable activity. The next step is building models that help defenders interpret that activity faster than adversaries can adapt.