If It Walks Like an Arbitrage: Protocol-Agnostic Detection with Decidable Structural Equivalence
This paper presents a protocol-agnostic, formally verified system that uses a decidable canonical form of Ethereum transaction traces to detect arbitrage opportunities via structural cycle detection, achieving high agreement with production MEV platforms while identifying thousands of exclusive detections without relying on protocol-specific patterns.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
On the Ethereum blockchain, money moves through a complex web of digital contracts. Every time a user sends a transaction, a hidden script runs, calling other programs in a nested sequence that can last for hundreds of steps. Within this invisible machinery, automated bots constantly scan for tiny price differences between digital exchanges. When they find one, they execute a rapid series of trades to buy low and sell high, capturing the profit before the market corrects itself. This activity, known as arbitrage, is the engine behind a massive industry called Maximal Extractable Value, where bots compete to squeeze every possible cent of profit from the network. However, because these bots use thousands of different contracts and constantly invent new ways to move money, spotting them has been like trying to find a specific needle in a haystack that keeps changing its shape. Existing tools often rely on memorizing the specific patterns of known exchanges, meaning they miss anything new or unusual.
A team of researchers at Functori has developed a new way to see these transactions that does not depend on knowing the names of the contracts involved. Instead of looking for specific patterns, they treat every transaction as a structural object that can be simplified into a single, unique shape. They built a system that takes the raw record of a transaction and rewrites it, step by step, stripping away the noise of the call hierarchy until only the essential flow of funds remains. This process is guaranteed to produce the same final shape every time, no matter how complex the original transaction was. If two different transactions result in the same final shape, they are structurally identical. This allows the system to identify arbitrage not by recognizing a known pattern, but by seeing if the money flows in a closed loop that starts and ends with more value than it began with. It is a method that works regardless of which digital tokens are used or which specific exchanges are involved.
The researchers tested this system on 220,000 blocks of the Ethereum blockchain, covering a month of activity in late 2025. They compared their results against Eigenphi, a leading commercial platform that detects arbitrage, and ArbiNet, a machine-learning classifier. The new system agreed with the commercial platform on 83.5% of the detected transactions, but it also found 60,199 confirmed arbitrage cycles that the commercial platform missed. These missed detections were often complex, multi-step trades involving flash loans or unusual routing that did not fit the standard patterns the commercial tool was trained to recognize. The system also identified 245,497 transactions that looked like arbitrage attempts but failed to make a profit, often because the bots lost a bidding war for transaction speed or paid too much in fees. These failed attempts are invisible to systems that only look for successful profits, but they reveal the true scale of the competition happening on the network.
To understand how this works, imagine the researchers taking a tangled ball of string representing a transaction and systematically untangling it. They do not care what the string is made of or where it came from; they only care about how the ends connect. If the string forms a loop where the end connects back to the start, and the loop contains a net gain of material, the system flags it as an arbitrage. This approach is so robust that the same software binary, without any changes, was able to run on two other blockchains, Arbitrum and Binance Smart Chain, and produce valid results. The system relies on a set of 15 rules that simplify the transaction tree, merging parallel paths and chaining transfers together until the structure either collapses into a simple cycle or reveals that no cycle exists. Because the rules are mathematically proven to always finish and always produce the same result, the detection is sound by construction.
The study also revealed that many transactions flagged as arbitrage by other systems were actually false positives. When the researchers manually inspected 200 transactions that only the commercial platform had flagged, they found that 63.5% of them did not contain any cyclic flow of funds at all. These were simple transfers or yield harvesting operations that looked like arbitrage only because the other system was looking for specific event signatures rather than the underlying structure. Conversely, the new system found genuine arbitrage cycles that other tools missed because they were hidden inside complex routing contracts. The researchers verified 500 transactions manually and found no false positives in their confirmed tier, meaning every transaction they labeled as a successful arbitrage was indeed a profitable loop.
This work moves the field from a reliance on memorized patterns to a fundamental understanding of value flow. By proving that every transaction has a unique, canonical form, the researchers have created a foundation where questions about strategy families and bot behavior can be answered with mathematical certainty. They demonstrated that arbitrage is not just a collection of known tricks, but a structural property of fund flows that can be detected without knowing the specific protocols involved. The system's ability to uncover hidden strategies and failed attempts suggests that the true volume of automated trading on the blockchain is significantly higher than previously measured, driven by a diverse ecosystem of bots that operate in ways that standard pattern-matching tools cannot see. The findings offer a new lens for forensic analysis, allowing investigators to trace the exact path of funds and understand the mechanics of profit extraction with a level of clarity that was previously impossible.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.