When Priority Fails:

Revert-Based MEV on Fast-Finality Rollups

Krzysztof M. Gogol1 , Manvir Schneider2 , and Claudio J. Tessone1,3

1
University of Zurich UZH
2
Cardano Foundation
3
UZH Blockchain Center
arXiv:2506.01462v4 [[Link]] 22 Sep 2025

Abstract. We study the economics of transaction reverts on Ethereum

rollups and show that they are not accidental failures but equilibrium
outcomes of MEV strategies. Using execution traces from major L2s, we
find that over 80% of reverted transactions are swaps, with half targeting
USDC–WETH pools on Uniswap v3/v4. Clustering reveals distinct bot
archetypes—including split-trade arbitrageurs, atomic duplicators, and
end-of-block spammers—demonstrating that reverts follow systematic
patterns rather than random noise. Empirically, we show that priority
fee auctions on rollups do not allocate blockspace efficiently: transac-
tion placement is mis-ordered, round-number bidding dominates, and
duplication spam inflates base fees. As a result, reverted transactions
contribute disproportionately more to sequencer fee revenues than to
gas consumption, shifting welfare from users to sequencers. To explain
these dynamics, we develop a model proving that trade-splitting and
duplication strictly dominate one-shot execution under convex adversar-
ial loss. Our findings establish reverts as a structural feature of rollup
MEV microstructure and highlight the need for protocol-level reforms to
sequencing, fee markets, and revert protection.

Keywords: MEV · Rollups · Sequencer Design · Priority Fee Auctions

1 Introduction
Maximum Extractable Value (MEV) [5] has evolved significantly since the early
days of Ethereum. The introduction of MEV Boost mechanisms, followed by
Proposer-Builder Separation (PBS), transformed how MEV was captured on
Ethereum, and now, with the rise of rollups [25], Layer-2 (L2) scaling solutions
for Ethereum, history appears to be repeating itself.
The centralization of sequencers significantly impacts MEV on rollups. A se-
quencer is the entity responsible for ordering transactions, forming blocks, and
submitting them to the Layer 1 chain [20]. As the sole actor capable of trans-
action ordering, the sequencer introduces implicit trust assumptions—namely,
that transactions are processed on a First-Come, First-Served (FCFS) basis with
Priority Fee Auctions (PFA). Moreover, most rollups (e.g., Arbitrum, Optimism,
Base, Unichain, ZKsync) operate private mempools. MEV opportunities on L2s
2 K. Gogol, M. Schneider, C. Tessone

are thus limited to back-running strategies such as arbitrage and liquidations,

or probabilistic (statistical) front-running or sandwiching [27]. Due to the low
latency of rollups (200 ms - 2 s), the extraction of MEV becomes a latency race
to be the first to submit transactions to the sequencer.
In March 2024, the Dencun upgrade [6] introduced blobs [7], a temporary
data storage mechanism designed to enhance rollup scalability. This upgrade
led to a significant reduction in gas fees on L2s, bringing them below $0.01,
[14], but it was accompanied by a noticeable increase in the transaction revert
rate [9]. While L2 revert rates oscillated below 5% prior to the upgrade, they
rose to over 10% afterward—by contrast, the revert rate on Ethereum mainnet
remains around 1%. These trends suggest the emergence of new MEV strategies
on rollups that might include submitting multiple copies of the same arbitrage
transaction to race for inclusion, splitting transactions into smaller parts, and
probabilistically targeting specific positions within the block (e.g., beginning or
end) to exploit sandwich or back-running opportunities, which we investigate in
details in this work.
Looking forward, reducing the rate of reverted transactions is essential for
L2s. Although blob transactions were initially cheap, their costs are expected to
rise, potentially impacting rollup scalability. Notably, reverted transactions are
stored on-chain and consume resources, despite not modifying the blockchain
state. To address these concerns, some rollups are exploring mitigation tech-
niques. For example, Unichain [1] is developing a revert protection mechanism
and MEV tax [22], whereas Arbitrum has introduced Time Boost [17], a mech-
anism that allows sequencers to prioritize MEV opportunities ahead of other
transactions. This paper contributes both empirically and theoretically to bet-
ter design and evaluations of such mechanisms.

Related Work. Zhu et al. [30] present a game-theoretic model of MEV auctions
under revert protection, showing theocratically that revert protection increases
auction revenue, market efficiency, and blockspace usage by incentivizing deter-
ministic participation. Our work empirically validates its assumptions.
The empirical study of MEV, especially on L2 blockchains, remains relatively
underexplored, and involves studies on non-atomic arbitrage [13,12,31] and cyclic
one [27]. Solmaz et al. [24] analyze cyclic arbitrage bots on L2s, termed opti-
mistic MEV, and find that they consume disproportionately more gas than they
contribute in fees. In contrast, we show that for the broader class of reverted
transactions, the reverse holds: sequencers earn disproportionately more from
fees than they expend in gas.
TimeBoost, introduced by Arbitrum (April 2025), is a transaction ordering
policy for rollup sequencers that incorporates both transaction timestamps and
bids [18]. Simulation results suggest that it can yield returns comparable to
or exceeding those of simple FCFS ordering for L2 MEV searchers [10]. Flash-
blocks, introduced on Base (July 2025) and Unichain (August 2025), provide
early transaction pre-confirmation every 200 ms by splitting each block into ten
 Revert-Based MEV on Fast-Finality Rollups 3

mini-blocks. Their effects on transaction ordering have not yet been empirically
assessed—a gap this work aims to fill.

Contributions. This paper makes three main contributions.

(i) Empirical. We provide the first systematic evidence that reverts on
Ethereum rollups are structured MEV outcomes rather than random waste. Over
80% are swap attempts, concentrated in USDC–WETH pools, generated by a
small number of identifiable bot archetypes. The persistence of these patterns
across Arbitrum, Base, Optimism, Unichain, and ZKsync shows that revert-
heavy equilibria are structural to fast-finality rollups, independent of sequencing
design.
(ii) Market inefficiency. We demonstrate empirically that priority fee auc-
tions on rollups are inefficient in practice. Transaction placement exhibits sys-
tematic mis-ordering, round-number bidding, and duplication spam, which in-
flate base fees for users while allowing sequencers to capture disproportionate
revenues from failed attempts relative to their gas consumption. Moreover, the
introduction of Flashblocks on Base and Unichain reduced global fee-ordering
efficiency, localizing competition to the first slot while leaving subsequent slots
largely insensitive to priority fees. These results challenge the assumption that
PFAs allocate blockspace efficiently under fast-finality conditions.
(iii) Theoretical. We develop a formal model that rationalizes the observed
strategies. We prove that trade-splitting and duplication strictly dominate single-
shot execution under convex adversarial loss, explaining the prevalence of frag-
mented swaps and high revert rates.
By linking empirical evidence with formal analysis, our work establishes re-
verts as equilibrium outcomes of rational MEV strategies and highlights the need
for protocol-level reforms to sequencing and revert protection on rollups.

2 Background on Rollups

According to the blockchain scalability trilemma [19], a blockchain can prior-

itize at most two of the following three properties: decentralization, security,
and scalability. Ethereum, the leading DeFi blockchain by TVL, prioritizes de-
centralization and security. This design choice has led to network congestion,
high gas fees, and limited throughput—approximately 12 transactions per sec-
ond (TPS). To address these limitations, scaling solutions have been developed
at both Layer 1 (L1) and Layer 2 (L2). L1 scaling introduces new blockchains
with alternative consensus mechanisms [16], sharding [28], and independent in-
frastructure. In contrast, L2 scaling executes intensive computations off-chain,
posting compressed results to the underlying L1 [23,11].

Rollups. Rollups [26] are non-custodial L2 solutions that function as indepen-

dent blockchains: they execute transactions, produce blocks, and periodically
submit compressed transaction data to the L1. This design allows rollups to
4 K. Gogol, M. Schneider, C. Tessone

inherit the security guarantees of the underlying L1—e.g., Ethereum’s staked

ETH—making it infeasible to tamper with rollup data without compromising
L1 security.
Optimistic rollups [15] assume transactions are valid unless challenged. This
assumption simplifies implementation and accelerates compatibility with the
Ethereum Virtual Machine (EVM), helping optimistic rollups become the first
to attract DeFi adoption. However, their fraud-proof mechanism introduces a
withdrawal delay, typically enforced through a 7-day challenge period.
ZK-rollups [4] use zero-knowledge proofs (ZKPs) to validate state transitions.
After a proof is generated off-chain by a prover, it is verified on-chain by a
smart contract (verifier ). This architecture ensures rapid finality and enhances
compression—only proofs, not raw transaction data, need to be posted to L1.
The trade-off is increased computational overhead due to proof generation.
A sequencer [3] is a key component of rollup infrastructure, responsible for
ordering transactions, forming blocks, and bundling them into batches submit-
ted to L1. This improves gas efficiency relative to L1 execution. Despite cryp-
tographic guarantees from optimistic and ZK rollups, sequencers are currently
centralized. Major rollups like Arbitrum, Optimism, Base, Unichain, ZKsync,
and StarkNet rely on centralized sequencers—and in the case of ZK-rollups,
centralized provers as well. Efforts are underway to decentralize both [21,29].

Rollup Rollup

Block Block ... Block Users Block Block ... Block Users

Sequencer Sequencer Batch Prover

Layer-1 Blockchain Layer-1 Blockchain

Commit Batch Commit Batch Batch Proof

Rollup Rollup
Smart Contract Smart Contract

Verifier
Block Block ... Block Block Block ... Block

(a) Optimistic rollup (b) ZK-rollup

Fig. 1: Most rollups operate with centralized sequencers that maintain private
mempools. These sequencers have exclusive control over transaction ordering
within the rollup.

Blockchain finality refers to the point at which a transaction becomes irre-

versible. On rollups, we distinguish between soft and hard finality. Soft final-
ity occurs when a transaction is accepted by the sequencer and included in an
L2 block, making it irreversible from the rollup’s perspective. Hard finality is
achieved once the corresponding batch is posted and confirmed on L1, securing
the transaction with L1 guarantees. As rollups are currently centralized, users,
 Revert-Based MEV on Fast-Finality Rollups 5

bridges, and MEV searchers typically act upon soft finality, which occurs every
200ms–2s—much faster than Ethereum’s 12s L1 block time.

Gas Fees on L2s. Although reverted transactions do not alter the blockchain
state, they are still included in the ledger and consume resources during ex-
ecution. In general, on Ethereum rollups, the total fee paid by a transaction
comprises two main components:
– L2 Execution Fee: the cost of executing the transaction on the rollup’s virtual
machine.
– L1 Data Availability Fee: the cost of posting transaction calldata to Ethereum
L1, which ensures data availability and security.
The L2 execution cost itself can be further decomposed into a base fee and a
priority fee, leading to the following structure:

Gas Fee on L2 = Base Fee + Priority Fee + L1 Data Fee

| {z }
L2 Execution Fee
| {z }
Gas Fee

These components can be computed from on-chain transaction data using the
following formulas [2]:

Execution Fee = effective gas price × gas used

Priority Fee = priority fee per gas × gas used
Base Fee = (effective gas price − priority fee per gas) × gas used

3 Methodology and Data

We analyze successful and reverted transactions on five EVM-compatible rollups,
Arbitrum, Optimism, Base, ZKsync, and Unichain, between January and July
2025, using Dune archive nodes as our data source. All analyzed rollups operate
private mempools under centralized sequencers, with block times between 0.2s
and 2s (vs. 12s on Ethereum). Table 1 summarizes their key parameters.

Data Collection. Users interact with blockchains through atomic transactions,

which revert entirely if contract conditions are violated (e.g., slippage limits in
swaps). Although reverted transactions do not emit event logs, their execution
steps remain visible in the Traces table. We collect transaction data from Dune’s
Transactions and Traces tables, and supplement successful transactions with
event data from EventLogs.

Swap Labeling. A transaction is labeled as a swap if it interacts with a DEX

pool or router and attempts token transfers. Labels from Dune’s [Link]
are extended manually to cover missing or newly deployed exchanges.
6 K. Gogol, M. Schneider, C. Tessone

Table 1: Technical characteristics of analyzed Ethereum rollups: L2s rely on pri-

vate mempools operated by centralized sequencers [31] with Priority Fee Auc-
tions (PFAs). Arbitrum introduced TimeBoost (TB) sequencing on April 2025,
Base introduced Flashblocks (FB) in July 2025, followed by Unichain in August
2025. The analyzed time period spans from January 2025 to August 2025.
Rollup Type Mempool Block Order Launch
Time Flow
Arbitrum (ARB) Optimistic Private 0.25s FCFS+TB Aug 21
Base (BASE) Optimistic Private 2.0s . PFA Aug 23
(0.20s FB)
Optimism (OP) Optimistic Private 2.0s PFA Dec 21
Unichain (UNI) Optimistic Private 2.0s . PFA Apr 24
(0.20s FB)
ZKsync (ZK) ZK Private 1.0s FCFS Mar 23

Atomic MEV Detection. We identify cyclic (atomic) MEV transactions following

the methodology of Torres [27] and Flashbots [8].

Graph Construction for Reverted Swaps. Reverted transactions are chal-

lenging to analyze because they do not emit application-layer event logs. The
EVM, however, still records their internal execution steps as execution traces.
We leverage this property to construct a call graph for each reverted transac-
tion. Formally, the execution graph of a transaction τ is a directed rooted tree
Gτ = (V, E), where each node v ∈ V corresponds to an internal execution
step (i.e., a trace entry), and each edge (vi → vj ) ∈ E denotes a call relation-
ship initiated by vi . Each node is annotated with its operation type (e.g., CALL,
DELEGATECALL, STATICCALL), caller and callee addresses, ETH value transferred,
calldata and returndata, and the execution status (success or revert).
To identify attempted DEX swaps, we traverse Gτ and match to addresses
against a canonical list of pools and routers (from Dune’s [Link], extended
manually). For Uniswap v2/v3 (and forks), we directly verify pool or router
calls with ERC-20 transfer selectors. In contrast, Uniswap v4 requires a differ-
ent approach: because all liquidity is mediated through a shared PoolManager
and hooks, we detect swaps by locating calls to the PoolManager, interac-
tions with known tokens (e.g., WETH, USDC), and calldata patterns consistent
with modifyPosition, swap, or hook-triggered functions. Transactions satisfying
these conditions are labeled as reverted swaps, and the corresponding subgraph
is extracted for further analysis.

4 Empirical Analysis of Reverted Transactions

 Revert-Based MEV on Fast-Finality Rollups 7

100.0%
Arbitrum Base Optimism Unichain ZKsync

Daily Transaction Revert Rate 80.0%

60.0%

40.0%

20.0%

0.0%
Jan 2025 Feb 2025 Mar 2025 Apr 2025 May 2025 Jun 2025 Jul 2025 Aug 2025

(a) Daily transaction revert rates on Ethereum L2s.

25.0% 100.0%

20.0%

Share of swaps in daily reverts

80.0%
15.0%
Diff in Revert Rate

60.0%
10.0%

5.0%
40.0%

0.0%
20.0%
−5.0%

−10.0% 0.0%
Arbitrum Base Optimism Unichain ZKsync Arbitrum Base Optimism Unichain ZKsync

(b) Difference in daily revert rates be- (c) Daily share of swaps among reverted
tween transactions that set priority fee transactions.
and all transactions.

Fig. 2: Analysis of transaction reverts from January to August 2025. (b) A posi-
tive difference on most L2s indicates that priority-fee transactions, often linked
to MEV, revert more frequently. (c) Reverts are predominantly swaps, reflecting
MEV/DEX activity.

Revert Rate Characteristics. Figure 2a shows that daily revert rates ranged from
5% to 40% across L2s, with peaks up to 20% on Arbitrum. On Unichain, revert
rates increased sharply in mid-May and remained close to 40% until July. Fig-
ure 2c demonstrates that most of reverted transactions are swaps. In particular,
more than 80% of reverted transactions are swaps on Arbitrum, Opti-
mism, and Unichain, suggesting their connection to arbitrage. Due to incomplete
labeling, these values represent lower bounds, especially on Base.

Attributes of Reverted Swaps. A detailed breakdown of reverted swaps charac-

teristics is presented in Figure 3. The majority of reverted swaps interact
with USDC–WETH pools and Uniswap v3, suggesting that most origi-
nate from MEV searchers targeting highly liquid pairs. On Arbitrum, Uniswap
v3 dominates with over 60% of reverted swaps. On Unichain, Uniswap v4 is the
primary target, accounting for more than 90% of all reverted swaps. Base ex-
hibits a more balanced distribution across Uniswap v2, v3, and v4 (each ∼20%).
8 K. Gogol, M. Schneider, C. Tessone

Arbitrum Base Optimism Unichain ZKsync

100%
97.1

80%
Top Platforms
60%
60.0

40%

32.1 32.0 31.5

20% 27.2
23.2 22.8 21.1
19.2 18.7 19.9 8.4
5.6 14.1 16.4
11.0 4.2 12.4 2.4 10.7
0.4 0.1 9.4
0%
Uniswap v3

Pancakeswap v3

Camelot v3

Uniswap v4

Others

Others

Uniswap v3

Uniswap v4

Uniswap v2

Pancakeswap v3

Uniswap v3

Velodrome v2_cl

Others

Velodrome v4

Uniswap v2

Uniswap v4

Uniswap v3

Velodrome v4

Velodrome v2_cl

Others

SyncSwap v1

Others

Maverick v1

ZKSwap_Finance v1

Mute v1
100

80
80.1
Top Token Pairs

60
52.8
40 42.5
39.4
33.7 30.9 29.5 28.8
20 25.3
21.7 21.7
14.9 5.4 5.2 5.3 5.2 7.8 7.0 7.8 5.0
12.1 4.0 3.1
10.0 0.8
0
US USOt rET WB Ot US w P A US O WB un aO WB U U O U US O U WE U
DC D he H- her DC eirdo SX-W RBU DC thers TC iBTC ptU TC SDC SD₮ thers NI-W DC thers SDT TH SDC
-W ₮0-W rs ws TC-W s -W - S-V -W -W S -W -W 0 ET .e- -W - -W
ET ET tET ET ET ZOR ETH IR ET ET -WB DCn- ET ET -WE H WE ET ZK ET
H H H H H A TU H H TC aO H H TH TH H H
AL ptW
BT
C

Fig. 3: Breakdown of reverted swaps by target DEX and token pair. Most re-
verted swaps target WETH–USDC pools and Uniswap (v3).

Arbitrum Base Optimism Unichain ZKsync

40k 3500 60k
800
3000 50k
# Reverted Transactions

6000
30k
2500 600
40k
2000
20k 4000 30k 400
1500
20k
2000 1000
10k 200
500 10k

0 0 0 0 0
50 100 150 100 200 300 400 20 40 60 80 10 20 30 40 50 0 5 10 15 20

Fig. 4: Index of reverted transactions within blocks. On Arbitrum and ZKsync,

reverts cluster at the block start; on Base, Optimism, and Unichain, they peak
around the fourth position.

The most frequently targeted pair is USDC–WETH (over 50% of reverts on

Arbitrum), with additional focus on USDC–USDT and WBTC–WETH.

Priority Fees and Auction Participation. Users may specify a priority fee per
gas to incentivize inclusion. Figure 8, in Appendix, shows the distribution of pri-
ority fees set by reverted transactions. Strikingly, nearly half of reverted trans-
actions—and almost all on ZKsync—include a priority fee despite the absence of
a priority fee auction on these L2s. On Unichain, nearly all reverted transactions
set the priority fee to exactly 1 wei, indicating minimal auction engagement. Fig-
ure 2b compares revert rates of priority-fee transactions with all transactions.
Base and Optimims exhibit a positive difference, indicating that transactions
setting priority fees revert more often. This counterintuitive result is ex-
plained by their association with MEV strategies, which we evaluate further.
 Revert-Based MEV on Fast-Finality Rollups 9

Arbitrum Base Optimism Unichain ZKsync

25

20
% from Reverted

15

10

0
Ga Ba Prio L1 To Ga Ba Prio L1 To Ga Ba Prio L1 To Ga Ba Prio L1 To Ga Ba Prio L1 To
s Us se rity Fe tal s Us se rity Fe tal sU se rity Fe tal sU se rity Fe tal sU se rity Fe tal
Fe Fe Fe Fe Fe Fe Fe Fe Fe Fe
ed e Fe e e ed e Fe e e sed e Fe e e sed e Fe e e sed e Fe e e
e e e e e

Fig. 5: Share of reverted transactions in gas usage and L2 fee revenue. Reverted
transactions contribute disproportionately more to fee revenue than to gas usage.

Block Positioning of Reverts. Figure 4 analyzes intra-block positioning. On Arbi-

trum and ZKsync, reverted transactions cluster at the very beginning of blocks,
consistent with intense low-latency races. On Base, Optimism, and Unichain, re-
verts peak later (around position 4), indicating different MEV execution strate-
gies under longer block times.
Figure 7 in Appendix C presents the relationship between the priority fee set
and the block index position of reverted transactions across L2s. On Arbitrum
and ZKsync, no clear pattern emerges (FCFS L2s). In contrast, on Base, Opti-
mism and Unichain we observe a consistent trend: transactions offering higher
priority fees are typically included closer to the beginning of the block. An addi-
tional notable observation is the concentration of reverted transactions around
“round-number” priority fees (e.g., 100 wei, 1 million wei). This suggests a strong
competition for inclusion at these discrete bidding levels and highlights the preva-
lence of suboptimal bidding strategies among MEV bots.
Appendix A analyzes the impact of Flashblocks on the efficiency of PFAs.
The results show that only the first Flashblock exhibits efficient fee-based
ordering, while in subsequent slots the priority fee has negligible influence on
transaction position within the block. This suggests that priority fee auctions
are not functioning as intended on fast-finality rollups, where latency
dominates over bidding.

Gas Usage and Fees. Figure 5 shows that reverted transactions contribute
disproportionately more to sequencer fee revenues than to gas usage.
Across all L2s, their share of total fees paid exceeds their share of gas consumed.
On Base and Unichain, reverted transactions generate ∼20–25% of priority fee
revenue. A longitudinal view of daily gas usage and fee contributions is provided
in Appendix C (Figure 9).

5 Address Clustering and MEV Strategies

We apply a clustering approach to identify MEV strategies and bots involved
in reverted transactions. We cluster blockchain addresses based on their
occurrence as the recipient (”to” address) in reverted transactions,
10 K. Gogol, M. Schneider, C. Tessone

since these addresses correspond to contracts that execute MEV strategies rather
than EOAs that merely submit them. Unlike senders, which are often one-off
spam shells, recipient contracts (e.g., routers, arbitrage bots, or swap helpers)
are stable points of interaction that generate internal calls (traces) to DEX
routers and liquidity pools, where reverts typically occur. Clustering at this
level allows us to identify recurring patterns of MEV activity, distinguish user
contracts from automated bots, and uncover characteristic patterns of failed or
repeated arbitrage attempts.
On May 20th, 2025, we identified 759, 1938, 578, 99, 68 such addresses on
Arbitrum, Base, Optimism, Unichain, and ZKsync, respectively. Due to the rel-
atively small sample sizes on Unichain and ZKsync, we focus our clustering
analysis on Arbitrum, Base, and Optimism.
For each one of those instances, we considered multiple factors, such as num-
ber of transactions, number of distinct addresses that called the analyzed ad-
dress, daily coverage (number of hours with at least one transaction), revert rate,
and found the following set of features significant:
1. Swap % - percentage of swaps among the successful transactions,
2. Atomic % - percentage of atomic arbitrage swaps among the successful trans-
actions,
3. Revert/Block - average number of reverted transactions in a block among
blocks with reverted transactions,
4. Success/Block - average number of successful transactions in a block among
blocks with successful transactions,
5. Revert No - number of reverted transaction in a day.
We perform Principal Component Analysis (PCA) on the standardized ver-
sion of our selected features. The first five principal components explain approxi-
mately {26.5%, 24.1%, 18.6%, 16.1%, 14.7%} of the variance, respectively. Hence,
the variance is distributed relatively evenly, and no single dominant component
captures the majority of the variability in the data.
We proceed to cluster our data via KMeans++ and experiment with different
seeds to initiate the algorithm with, and with the possible number of target
clusters to look for (we test for values ranging from 2 to 10). The related inertia
plot is shown in Figure 6a, which clearly suggests 6 as the optimal number of
clusters to look for. We therefore save the KMeans++ clustering obtained for
such 6 groups, and plot the t-SNE projection of our points in Figure 6b, where
the points are color-coded by the KMeans++ clustering labels.
Figure 6 summarizes the features for each cluster, in order to help with
their interpretation and discussion. For each group, we report the number of
points belonging to it, but also show the average value of each feature that was
considered during our clustering phase. In general, we can make the following
observations.

Group 0: Passive Receivers. This is the largest cluster, containing the majority
of recipient contracts. These addresses exhibit almost no interaction with swaps
(swap% ≈ 0), very low atomic arbitrage involvement, and only a small number
 Revert-Based MEV on Fast-Finality Rollups 11

(a) Inertia plot pointing to optimal clus-

ters. (b) t-SNE projection of addresses.

# of Swap Atomic Reverts Success Revert Reverts

Gr. Name Recipients % % / Block / Block Rate / Day
0 Passive Receivers 1630 0.0% 0.0% 1.00 1.00 4.0% 2
1 Single-Shot Arbitrageurs 1032 100.0% 0.0% 1.00 1.00 25.0% 6
2 Tx Relays / Spam Shells 1 0.0% 0.0% 36.90 85.30 19.9% 4,135
3 Split-Trade Arbitrageurs 2 72.2% 0.0% 4.40 3.05 58.2% 150,894
4 End-of-Block Spammers 2 50.0% 0.0% 196.65 2.95 51.0% 27,937
5 Atomic Duplicators 68 100.0% 67.2% 1.00 1.00 46.8% 98

Fig. 6: We cluster our data points relating to reverted transactions. The inertia
plot suggests an optimal clustering into 6 groups, visualised via a t-SNE projec-
tion and summarised in the table.

of reverted transactions per block. Their activity is sporadic, with limited daily
coverage, suggesting that they correspond to non-MEV infrastructure contracts
or passive user contracts that occasionally receive reverted calls, rather than
specialized arbitrage bots. In other words, Group 0 forms the background noise
of the reverted transaction landscape.
Example. The contract 0x42d...9a6 on Base illustrates this passive-receiver
pattern, with rare reverted calls but no consistent swap activity.

Group 1: Single-Shot Arbitrageurs. Addresses in this group almost exclusively

execute swaps (swap% close to 1), with on average one successful swap and one
revert per block. The strategy is binary: either the arbitrage succeeds or the
transaction reverts. This pattern is consistent with a single-shot CEX–DEX
arbitrage strategy, where arbitrageurs attempt to capture price discrepancies
between a centralized exchange and a single DEX pool with one atomic swap.
When the opportunity vanishes during block inclusion, the trade simply fails.
Example. The address 0x436d...795 on Optimism repeatedly engages Uniswap
v3 pools in this binary success-or-revert pattern.

Group 2: Transaction Relays / Spam Shells. This cluster contains only a few
addresses, but with extremely high numbers of both reverted and successful
transactions per block. They do not rely on swaps (swap% ≈ 0), indicating that
these may be auxiliary contracts or router helpers that aggregate or forward
12 K. Gogol, M. Schneider, C. Tessone

bundles of calls. Their high throughput per block suggests that they act as
transaction relays or spam shells, rather than genuine arbitrage bots.
Example. The contract 0xc40d...f90 on Base shows this pattern of dense trans-
action forwarding without swap involvement.

Group 3: Split-Trade Arbitrageurs. This group produces the largest number of

reverted transactions per day, with revert rates exceeding 50%. It contains only a
handful of addresses, but each shows a distinct pattern: transactions are swaps,
and if successful they appear multiple times within the same block, otherwise
they revert. This behavior is consistent with a CEX–DEX arbitrage strategy
with trade splitting, where a larger arbitrage opportunity is decomposed into
several smaller swap attempts to mitigate slippage and latency risk. Such split-
ting maximizes expected returns but naturally generates many failed attempts,
explaining the observed dominance in reverted transactions.
Examples. 0xb7b...463 on Base targets Uniswap v3 USDC–WETH swaps ex-
clusively, while 0xcb4...95c on Arbitrum splits trades across multiple pools
(USDC–WETH, USDT–WETH, WBTC–WETH, USDC–ARB), both exempli-
fying this split-arbitrage style.

Group 4: End-of-Block Spammers. This cluster is characterized by an unusually

high number of reverts per block combined with very few successful swaps. The
addresses appear to send repeated transactions that fail more often than not,
with activity peaking at end-of-block positions. This pattern fits a spam-based,
end-of-block sniping strategy, in which bots flood the sequencer with revert-
prone transactions in hopes that one lands in a favorable position to exploit
short-lived price dislocations caused by noise traders.
Examples. Two Base addresses illustrate this strategy: 0x730...bfb and 0x055...298,
both sending repeated spam-like swaps with high revert rates.

Group 5: Atomic Duplicators. Although relatively small in size, this cluster

stands out with very high swap% and atomic% values, and a revert rate above 50%.
Addresses here engage almost exclusively in arbitrage via swaps, often across
known DEX routers, and operate around the clock (timezone coverage ≈ 24).
The strategy resembles atomic arbitrage with duplication, where multiple
parallel attempts of the same swap are fired within a block to maximize the
chance of capturing a fleeting price gap. Compared to Group 1, these bots are
more aggressive and maintain continuous coverage of opportunities across time
zones.
Example. The Optimism address 0x4479...146 exemplifies this aggressive, around-
the-clock atomic duplication behavior.

6 Theoretical Model of Trade Splitting and Duplication

The prevalence of duplicated swaps and fragmented trade attempts observed

in Section 4 and 5 suggests that arbitrageurs strategically split transactions
 Revert-Based MEV on Fast-Finality Rollups 13

to optimize their chances of successful execution under slippage and latency

risks. Our theoretical model captures these dynamics, providing a foundation
for interpreting why such patterns of reverts and clustering emerge in practice.
It abstracts away from priority fees. In particular, priority fees they are absent
(e.g., Arbitrum, zkSync) or empirically negligible (e.g., Unichain), allowing us
to study duplication and splitting without loss of generality.
Token pair. There is a single token pair (X, Y ) traded on
– a constant-product AMM with reserves (x, y) and fee f ∈ [0, 1);
– a centralized exchange (CEX) that quotes a fixed price Pc > 0 expressed
in Y per X and always settles.
AMM swap mechanics. If an arbitrageur swaps q > 0 units of X with the AMM
liquidity pool, she receives
y(1 − f )q
∆y(q) = , (1)
x + (1 − f )q
and the reserves in the AMM liquidity pool update to (x + (1 − f )q, y − ∆y(q)).
Formula (1) is the standard CPMM payout.
Costs.
– A fixed per-swap overhead cg ≥ 0 (gas & latency).
– If the swap fails, the arbitrageur is left long q units of X; liquidating this
inventory incurs a penalty ϕ ≥ 0 (possibly 0).
Arbitrage objective. The trader wishes to arbitrage a total size D > 0 (in token X)
by

1. buying D units of X on the CEX for Pc D units of Y ;

2. selling those units into the AMM, possibly split into n ∈ N equal chunks of
size q = D/n, or send multiple copies of the full trade.

Per–swap success. For one attempted swap of size q, we have the following no-
tation.
– Success (p(q)): receive ∆y(q) from AMM, offset the Pc q spent on CEX.
– Failure (1 − p(q)): CEX trade clears, AMM trade reverts; trader still paid
Pc q on CEX and may liquidate at cost ϕ.
– Overhead cg is incurred regardless.
Adversarial loss. We model on-chain execution hazards (e.g., sandwiching) via
an adversarial loss function L : [0, ∞) → R+ applied to the submitted size q.
Economically, L(q) captures expected value extracted by adversaries around the
trade. In CFMMs, curvature of the pricing curve together with user slippage
bounds implies that the attackable “wedge” scales more than linearly with size,
so it is natural to assume:
Assumption 01 (Convex adversarial loss) L is nondecreasing, strictly con-
vex, and L(0) = 0.
14 K. Gogol, M. Schneider, C. Tessone

This abstraction bundles sandwich/visibility costs and similar effects; constants

and modeling details enter only through the shape of L(·).
Conditioning convention across subsections. There are two sensible ways to place
L relative to success probability s: (i) conditional L (loss realized only if the
swap executes), and (ii) unconditional L (already averaged over success). To
avoid double counting while keeping formulas simple, we adopt the following
convention:
– Section 6.1 (single-attempt splitting). Here the per-attempt success prob-
ability s is treated as a constant. We therefore absorb s into L and, with
slight abuse of notation, write L(·) for the expected adversarial loss per at-
tempt (already scaled by s). This preserves convexity and keeps statements
compact.
– Section 6.2 (duplicate submissions). Here the success probability depends on
the number of copies K (namely 1 − (1 − s)K ). To prevent spurious scaling
of loss with K, we keep L(·) conditional on success and multiply it by the
overall success probability when computing expectations.

6.1 Splitting trades

We first show that under convex adversarial loss, arbitrageurs can strictly im-
prove their expected profit by splitting a large trade into smaller slices.
Proposition 1 (Splitting under convex adversarial loss). Fix a CFMM
pool and a total input size D > 0. Let the per-attempt success probability be
constant p(q) ≡ s ∈ (0, 1]. Let L : [0, ∞) → R+ satisfy Assumption 01. Let
the per-attempt fixed overhead κ = cg + (1 − s)ϕ, (κ ≥ 0). Denote by ∆y(q)
the CFMM output of swapping size q from the initial state, and by {∆yi }ni=1 the
incremental outputs of executing n equal slices q = D/n back-to-back in the same
pool without any external state refresh. Define the expected profit from a single
attempt of size q as

π(q) = s∆y(q) − Pc q − L(q) − κ,

and the total expected profit from splitting into n equal slices as
n 
X 
Π(n) = s∆yi − Pc D/n − L(D/n) − κ .
i=1

Then
n
!
X
Π(n) − Π(1) = s ∆yi − ∆y(D) + L(D) − nL(D/n) − (n − 1)κ. (2)
i=1
Pn
In particular, under same-pool, same-block execution, i=1 ∆yi = ∆y(D) and


Π(n) − Π(1) = L(D) − nL(D/n) − (n − 1)κ. (3)
 Revert-Based MEV on Fast-Finality Rollups 15

Hence splitting strictly improves expected profit (relative to one big trade) when-
ever
L(D) − nL(D/n) > (n − 1)κ.4

Remark 1 (Conditioning of the adversarial loss L). In CFMM sandwiches the

attacker’s profit is realized only if the victim’s swap executes. Since in this section
the per-attempt success probability s is treated as constant, we absorb it into L
and interpret L(·) as the expected adversarial loss per attempt (already scaled by
s). This preserves convexity and leaves Proposition 1 unchanged (only coefficients
rescale, e.g., in the quadratic case).

Corollary 1 (Quadratic adversarial loss). If L(q) = αq 2 with α > 0, then

 
1
Π(n) − Π(1) = αD2 1 − − (n − 1)κ.
n

Treating n as a positive real and differentiating,

 
d   2 1
Π(n) − Π(1) = αD − κ.
dn n2
p
Setting this to zero gives the continuous optimum n⋆cont = D α/κ. The optimal
integer choice is
n  o
n⋆ = max 1, n⋆cont ,

i.e., the nearest integer.

6.2 Duplicate-Submission Spam

Fix a slice size q > 0 and let s := p(q) ∈ [0, 1] be the success probability of a
single submission of size q. If a trader submits K ∈ N identical copies such that
at most the earliest one can execute, then the success event occurs iff at least one
copy lands. Under the standard independent-arrival approximation, the success
probability is

K
sK (q) = 1 − 1 − s .

We model adversarial extraction (e.g., sandwiching) by a loss function L(q)

(satisfying Assumption 01) that is conditional on success: if one copy executes,
the realized gross benefit ∆y(q) is reduced by L(q), and if no copy executes, no
such adversarial loss is realized.5
4
Proof in Appendix B
5
In Section 6.1 we treated s as a constant per attempt and therefore absorbed it into
L without loss of generality. In duplication, the success probability depends on K via
1−(1−s)K , so keeping L conditional on success avoids artificially scaling adversarial
loss with K. See Remark 2 below.
16 K. Gogol, M. Schneider, C. Tessone

Let W := ∆y(q) + ϕ. The expected utility of choosing K copies is

U (K; q) = 1 − (1 − s)K ∆y(q) + ϕ − L(q) −Pc q − ϕ − Kcg
 

| {z }
success term
(4)
= 1 − (1 − s)K W − L(q) − Kcg −Pc q − ϕ.
 

| {z }
FL (K; q)
The term Pc q + ϕ does not affect the optimizer in K, so we maximize FL (K; q).
We next establish conditions under which duplicating the same trade multiple
times is optimal, and derive the formula for the optimal number of copies.
Proposition 2 (Optimal duplication). Fix q > 0 and define A := W −
L(q) = ∆y(q) + ϕ − L(q). Then:
(i) No duplication. If A ≤ 0 or sA ≤ cg , it is optimal to send a single copy:
K ⋆ (q) = 1.
(ii) Optimal duplication. If A > 0 and sA > cg , the optimal number of copies
is the largest integer K such that
FL (K; q) − FL (K − 1; q) = (1 − s)K−1 sA − cg ≥ 0.
Equivalently, $
%
ln cg /(sA)
K ⋆ (q) = 1 + 6
(5)
ln(1 − s)
Remark 2 (Why L is conditional here). Only the earliest-included copy can ex-
ecute; all other copies revert. Adversarial extraction (e.g., sandwich attacks)
requires the victim trade to execute. Hence L is naturally specified conditional
on success in the duplication problem, and it is multiplied by the total success
probability 1 − (1 − s)K in Equation (4). By contrast, in Section 6.1 we ana-
lyzed a single attempt with constant success probability s; there, absorbing s
into L is without loss of generality and simplifies the exposition. Keeping L con-
ditional here prevents scaling of adversarial loss with the number of duplicate
submissions.

7 Discussion
Reverts as Equilibrium. Our results establish that reverted transactions on
rollups are not accidental noise but equilibrium outcomes of rational MEV strate-
gies. Splitting and duplication strictly dominate single-shot execution under con-
vex adversarial loss, and empirically we observe exactly these patterns across L2s.
High revert rates are therefore an endogenous response to the interaction of fast
block times, private mempools, and fee rules, rather than evidence of inefficient
or malicious actors alone. Interpreting reverts as equilibrium behavior reframes
them from wasted blockspace to a predictable by-product of current sequencing
designs, thus reverts are best understood as equilibrium outcomes shaped by
sequencer policy, not anomalies.
6
Proof in Appendix B.
 Revert-Based MEV on Fast-Finality Rollups 17

Sequencer Mechanisms: Flashblocks and TimeBoost. The introduction of Flash-

blocks on Base and Unichain, and TimeBoost on Arbitrum, illustrates how small
changes in sequencer policy reshape rather than eliminate MEV equilibria (Ap-
pendxiA). Flashblocks concentrate fee competition in Slot 1, where revert proba-
bility rises sharply, while subsequent slots remain insensitive to fees. TimeBoost
introduces fee sensitivity at the block head, but FCFS ordering persists else-
where. In both cases, revert-heavy equilibria shift to new focal points rather
than disappear. These designs redistribute congestion costs but do not alleviate
the overall welfare loss for users.

Implications for Revert Protection. Reverted transactions contribute dispropor-

tionately to sequencer revenues—up to one-quarter of all priority fees on some
rollups—despite consuming far less gas. This asymmetry raises the question of
whether reverted transactions should be exempt from base and L1 data fees,
while still paying priority fees to compensate sequencers for processing. Such a
design could reduce welfare losses for ordinary users, but care is required: trans-
actions that currently settle without altering state could be strategically adapted
to qualify as reverts, creating new avenues for exploitation. The challenge is thus
to design revert protection that lowers systemic waste while preventing gaming
by sophisticated bots. Hybrid fee schemes that penalize duplication without
punishing isolated reverts may be a promising direction.

Cross-Rollup Generality. The persistence of inefficiencies across Arbitrum, Base,

Optimism, Unichain, and ZKsync suggests that revert-heavy equilibria are an
inherent risk of fast-finality rollups, rather than an artifact of a particular se-
quencing mechanism.

8 Conclusions

This paper examined the economics of reverted transactions on Ethereum rollups

through a combination of empirical measurement, clustering, and theoretical
modeling. We showed that reverts are not random noise but systematic outcomes
of MEV strategies: more than 80% are swap attempts, heavily concentrated in
USDC–WETH pools, and generated by identifiable bot archetypes such as split-
trade arbitrageurs, atomic duplicators, and end-of-block spammers. Our analysis
further revealed that priority fee auctions on rollups fail to allocate blockspace
efficiently. Mis-ordering, round-number bidding, and duplication spam inflate
base fees for users while enabling sequencers to capture disproportionate revenues
from failed attempts relative to their gas consumption.
To rationalize these dynamics, we developed a formal model proving that
trade-splitting and duplication strictly dominate single-shot execution under
convex adversarial loss, thereby explaining the prevalence of fragmented swaps
and high revert rates. Together, these findings reframe reverts as equilibrium
features of current rollup fee markets rather than wasteful anomalies.
18 K. Gogol, M. Schneider, C. Tessone

The implications are clear: without protocol-level reforms to sequencing, re-

vert protection, and transaction fee mechanisms, duplication-heavy equilibria
will persist, leaving rollup fee markets both inefficient and welfare-distorting.

Acknowledgments
The first author gratefully acknowledges the support of the Flashbots Research
Grant. The authors are also indebted to Danning Sui, Daniel Marzec, Johnnatan
Messias, and Xin Wan for their valuable discussions and insights, which greatly
contributed to shaping this research.

References
1. Adams, H., Toda, M., Karys, A., Wan, X., Gretzke, D., Zhong, E., Wong, Z.,
Marzec, D., Miller, R., Hasu, Floersch, K., Robinson, D.: Unichain (2024)
2. Buterin, V., Conner, E., Dudley, R., Slipper, M., Norden, I., Bakhta, A.: EIP-
1559: Fee market change for ETH 1.0 chain. [Link]
eip-1559 (2021), ethereum Improvement Proposal
3. Chaliasos, S., Firsov, D., Livshits, B.: Towards a Formal Foundation for Blockchain
Rollups. arXiv preprint arXiv:2406.16219 (2024)
4. Chaliasos, S., Reif, I., Torralba-Agell, A., Ernstberger, J., Kattis, A., Livshits, B.:
Analyzing and Benchmarking ZK-Rollups (2024)
5. Daian, P., Goldfeder, S., Kell, T., Li, Y., Zhao, X., Bentov, I., Breidenbach, L.,
Juels, A.: Flash Boys 2.0: Frontrunning in Decentralized Exchanges, Miner Ex-
tractable Value, and Consensus Instability. In: 2020 IEEE Symposium on Security
and Privacy (SP). pp. 910–927. IEEE (2020)
6. Ethereum Foundation: Ethereum Roadmap (2024), [Link]
roadmap/, accessed on February 7, 2025
7. Ethereum Foundation: Proto-Danksharding FAQ. [Link]
@vbuterin/proto_danksharding_faq (2024), accessed on Febraury 7, 2025
8. Flashbots: Atomic arbitrage. [Link]
(2024), accessed: August 4, 2025
9. Flashbots: Flashbots – Multichain Transaction Reverts. [Link]
flashbots/multichain-transaction-reverts (2024), accessed on February 7,
2025
10. Fritsch, R., Silva, M.I., Mamageishvili, A., Livshits, B., Felten, E.W.: Mev capture
through time-advantaged arbitrage (2025), [Link]
11. Gangwal, A., Gangavalli, H.R., Thirupathi, A.: A Survey of Layer-Two Blockchain
Protocols. Journal of Network and Computer Applications, Vol 209 (2022)
12. Gogol, K., Messias, J., Miori, D., Tessone, C., Livshits, B.: Cross-Rollup MEV:
Non-Atomic Arbitrage Across L2 Blockchains. arXiv preprint arXiv:2406.02172
(2024)
13. Gogol, K., Messias, J., Miori, D., Tessone, C., Livshits, B.: Quantifying Arbitrage in
Automated Market Makers: An Empirical Study of Ethereum ZK Rollups. In: The
5th International Conference on Mathematical Research for Blockchain Economy
(MARBLE) (2024)
14. Growthepie: Layer-2 Transaction Costs. [Link]
fundamentals/transaction-costs (2024), accessed on February 7, 2025
 Revert-Based MEV on Fast-Finality Rollups 19

15. Kalodner, H., Goldfeder, S., Chen, X., Weinberg, S.M., Felten, E.W.: Arbitrum:
scalable, private smart contracts. In: Proceedings of the 27th USENIX Conference
on Security Symposium (2018)
16. Lashkari, B., Musilek, P.: A Comprehensive Review of Blockchain Consensus Mech-
anisms. IEEE Access (2021)
17. Mamageishvili, A., Kelkar, M., Schlegel, J.C., Felten, E.W.: Buying Time: La-
tency Racing vs. Bidding in Transaction Ordering. arXiv preprint arXiv:2306.02179
(2023)
18. Mamageishvili, A., Kelkar, M., Schlegel, J.C., Felten, E.W.: Buying time: Latency
racing vs. bidding in transaction ordering (2023), [Link]
02179
19. Monte, G.D., Pennino, D., Pizzonia, M.: Scaling blockchains without giving up
decentralization and security. In: Proceedings of the Workshop on Cryptocurrencies
and Blockchains for Distributed Systems (2020)
20. Motepalli, S., Freitas, L., Livshits, B.: Sok: Decentralized sequencers for rollups
(2023), [Link]
21. Motepalli, S., Freitas, L., Livshits, B.: SoK: Decentralized Sequencers for Rollups.
arXiv preprint arXiv:2310.03616 (2023)
22. Robinson, D., White, D.: Priority Is All You Need. Paradigm Reseach (2024)
23. Sguanci, C., Spatafora, R., Vergani, A.M.: Layer 2 blockchain scaling: a survey.
arXiv preprint arXiv:2107.10881 (2021)
24. Solmaz, O., Heimbach, L., Vonlanthen, Y., Wattenhofer, R.: Optimistic mev in
ethereum layer 2s: Why blockspace is always in demand (2025), [Link]
org/abs/2506.14768
25. Thibault, L.T., Sarry, T., Hafid, A.S.: Blockchain Scaling Using Rollups: A Com-
prehensive Survey. IEEE Access, Vol. 10 (2022)
26. Thibault, L.T., Sarry, T., Hafid, A.S.: Blockchain Scaling Using Rollups: A Com-
prehensive Survey. IEEE Access (2022)
27. Torres, C.F., Mamuti, A., Weintraub, B., Nita-Rotaru, C., Shinde, S.: Rolling in
the Shadows: Analyzing the Extraction of MEV Across Layer-2 Rollups. arXiv
preprint arXiv:2405.00138 (2024)
28. Wang, G., Shi, Z.J., Nixon, M., Han, S.: SoK: Sharding on Blockchain. In: Pro-
ceedings of the 1st ACM Conference on Advances in Financial Technologies (2019)
29. Wang, W., Zhou, L., Yaish, A., Zhang, F., Fisch, B., Livshits, B.: Mechanism
Design for ZK-Rollup Prover Markets. arXiv preprint arXiv:2404.06495 (2024)
30. Zhu, B.Z., Wan, X., Moallemi, C.C., Robinson, D., Bachu, B.: Quantifying the
Value of Revert Protection. arXiv prepritnt arXiv:2410.19106 (2024)
31. Öz, B., Torres, C.F., Gebele, J., Rezabek, F., Mazorra, B., Matthes, F.: Pandora’s
Box: Cross-Chain Arbitrages in the Realm of Blockchain Interoperability. arXiv
preprint arXiv:2501.17335 (2025)

A Impact of Flashblocks and TimeBoost on Transaction

Ordering by Sequencer

Flashblocks, a form of early transaction pre-conformation, were introduced on

Base in July 2025 and on Unichain in August 2025. Both L2s, developed with the
OP Stack operate with block times of approximately 2 seconds, and Flashblocks
offer the pre-confiramtion every 200mn,
20 K. Gogol, M. Schneider, C. Tessone

Arbitrum sequencer blocks are produced at an interval of ∼200–250ms, pro-

viding substantially faster inclusion latency than OP Stack chains prior to the
introduction of Flashblocks. However, Arbitrum sequencer does not consider pri-
ority fee, when ordering the transaction (FCFS rule). Thus, in April 2025, Ar-
bitrum introduced TimeBoost, a modification of the sequencer ordering policy
that establishes a priority lane for the boosted transactions.

A.1 Mechanism Overview

Flashblocks. Flashblocks subdivide each L2 block into N smaller units, typically

N = 10, referred to as preconfirmation slots. Transactions are scheduled into
these slots sequentially, providing users with a low-latency guarantee (precon-
firmation) of inclusion before the full block is sealed. The design intention is
twofold: (i) to reduce latency by offering preconfirmations every ∼200ms rather
than once per block, and (ii) to create predictable inclusion windows.

TimeBoost (Arbitrum). TimeBoost augments the sequencing infrastructure with

two components: a sealed-bid second-price auction and a dedicated express lane.
Valid transactions submitted to the express lane are sequenced immediately with
no delay, while all other transactions experience a nominal deferment (default:
200ms). The auction winner is granted exclusive rights to control the express lane
for predefined, temporary intervals, enabling fee-based competition for ultra-low-
latency inclusion. The default block time for Arbitrum remains at ∼250ms, but
under TimeBoost some transactions outside the express lane may be postponed
to the subsequent block.

A.2 Methodology

Transaction ordering is normalized to the unit interval [0, 1], where 0 denotes
the block head and 1 the block tail. On Flashblocks-enabled chains, each block
is heuristically divided into ten equal slots to approximate sub-block ordering.
Priority fees are grouped into logarithmically spaced bins, with an additional
bin reserved for zero-fee transactions. For each fee bin and slot combination, we
compute the median block position and the share of reverted transactions, and
estimate regressions of median position on the log of the priority fee.

A.3 Flashblocks: Results on Base and Unichain

Before vs. After Heatmaps. Heatmaps of revert share (% of transactions re-

verted) across fee levels and slot position reveal a structural break. Prior to Flash-
blocks, revert clustering shows no discernible slot pattern, with mis-ordering
dispersed throughout the block. After introduction, reverts concentrate at the
block head (Slot 1), consistent with an auction-like mechanism.
 Revert-Based MEV on Fast-Finality Rollups 21

Table 2: Comparison of Flashblocks and TimeBoost as Sequencer Ordering

Mechanisms.
Feature Flashblocks TimeBoost

Launch Base: July 2025 Arbitrum: April 2025

Unichain: August 2025

Underlying stack OP Stack (Base, Unichain) Arbitrum Nitro stack

Block time (base- ∼2s before Flashblocks ∼200–250ms

line)

Structure Block subdivided into N preconfir- Single block structure maintained; intro-
mation slots (typically N = 10), duces a new express lane at the sequencer.
each with ∼200ms latency guaran-
tees.

Auction type Implicit priority fee competition Sealed-bid second-price auction deter-
within each slot. mines control of the express lane.

Inclusion rule Transactions placed sequentially Winning bidder’s transactions in the ex-
into slots; higher priority fees com- press lane are sequenced immediately; oth-
pete for earlier position in slots ers incur nominal delay (∼200ms) and may
be deferred to the next block.

Zero vs. Positive-Fee Comparison. Line plots of revert probability conditioned

on fee regime show that zero-fee transactions remain stable across both periods,
with consistently low revert shares. In contrast, positive-fee transactions are
disproportionately exposed to reverts after Flashblocks, particularly in Slot 1.

Median Position vs. Fee (Slot-by-Slot). Slot-level median positions against fee
reveal heterogeneous effects. Slot 1 exhibits a steep negative slope: higher fees re-
liably push transactions to the block head, but with elevated revert risk. Slots 2–9
display nearly flat relationships, indicating that fee has negligible ordering im-
pact. Slot 10 shows a weaker but persistent secondary effect: transactions “tail-
gate” behind Slot 1 and inherit some revert exposure.

Regression Analysis. Slot-specific regressions of transaction position on log pri-

ority fee (Table 3) confirm the descriptive patterns. For Base, Slot 1 exhibits
the steepest negative slope, consistent with a functioning auction at the block
head, while Slots 2–9 coefficients are near zero. Slot 10 shows a modest slope,
consistent with secondary clustering. Unichain displays a similar structure: sig-
nificant fee sensitivity in Slots 1, 6, 7, and 9, with other slots flat. These results
indicate that Flashblocks localizes fee competition to a subset of slots rather
than enforcing global ordering efficiency across the block.
22 K. Gogol, M. Schneider, C. Tessone

Arbitrum (pre-TimeBoost) Arbitrum (TimeBoost)

1 1

Tx Index Percentile
0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
1 100 10k 1M 100M 1 100 10k 1M 100M

Base (pre-Flashblocks) Base (Flashblocks)

1 1
Tx Index Percentile

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
1 100 10k 1M 100M 1 100 10k 1M 100M

Unichain (pre-Flashblocks) Unichain (Flashblocks)

1 1
Tx Index Percentile

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
1 100 10k 1M 100M 1 100 10k 1M 100M

Optimism ZKsync
1 1
Tx Index Percentile

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
1 100 10k 1M 100M 1 100 10k 1M 100M

Priority Fee per Gas (wei) Priority Fee per Gas (wei)

Median (Successful) Median (Reverted)

Fig. 7: Median priority fee across transaction index percentiles, with interquar-
tile range (IQR) bands, for each analyzed L2 chain. On Base, Optimism, and
Unichain, higher priority fees are associated with earlier inclusion in the block
(lower index positions), whereas no such pattern is observed on Arbitrum and
ZKsync.
 Revert-Based MEV on Fast-Finality Rollups 23

A.4 TimeBoost: Results on Arbitrum

In April 2024, Arbitrum introduced TimeBoost, a sequencing mechanism that
prioritizes transactions submitted with valid priority bids. We apply the same
empirical methodology as for Flashblocks: (i) heatmaps of revert share across fee
levels and slots, (ii) zero vs. positive-fee comparisons, (iii) median transaction
position vs. fee by slot, and (iv) regression analysis of fee–position sensitivity.

Before vs. After Heatmaps. Prior to TimeBoost, revert clustering exhibited no

strong slot structure, with transactions dispersed across the block irrespective of
fee. After TimeBoost, concentration at early positions (Slot 1) becomes evident,
indicating auction-like prioritization.

Zero vs. Positive-Fee Comparison. Zero-fee transactions remain largely unaf-

fected by TimeBoost, showing consistently low revert rates. By contrast, positive-
fee transactions display increased fee sensitivity post-TimeBoost: higher fees are
more strongly associated with early placement but also with elevated revert
probability at the block head.

Median Position vs. Fee (Slot-by-Slot). Slot-level analysis reveals that prior to
TimeBoost, the median position is only weakly responsive to fees. After Time-
Boost, Slot 1 shows a sharp negative slope (higher fees reliably move transactions
to the head), while Slots 2–9 remain flat, and Slot 10 displays a weaker tailgating
effect, similar to Base and Unichain under Flashblocks.

Regression Analysis. Pooled regressions of transaction position on log priority

fee (Table 4) show no statistically significant relationship either before or after
TimeBoost. This confirms that Arbitrum’s sequencer continues to operate un-
der a FCFS rather than a PFA for majority of transactions. While TimeBoost
introduces an auction-like mechanism in principle, its effects are not observable
in fee–position sensitivity at the block level, in sharp contrast to the localized
ordering observed under Flashblocks.

A.5 Discussion and Implications

Flashblocks. Across Base and Unichain, the introduction of Flashblocks reshapes
fee–ordering dynamics. The system approximates an auction only in Slot 1, while
Slots 2–9 show no fee discrimination. Reverts disproportionately cluster at the
block head, magnifying costs for participants bidding for early inclusion. Zero-
fee transactions, by contrast, are stable and rarely reverted, highlighting the
asymmetric burden of Flashblocks on fee-paying users.

TimeBoost. TimeBoost introduces meaningful fee sensitivity at the front of the

block, aligning Arbitrum’s sequencing closer to a priority auction. However, or-
dering efficiency is limited: only Slot 1 reflects strong auction dynamics, while
the majority of slots remain insensitive to fee bids. Reverts cluster more heavily
at the block head post-TimeBoost, indicating that competitive bidding for early
inclusion amplifies revert-based MEV strategies rather than eliminating them.
24 K. Gogol, M. Schneider, C. Tessone

Chain Slot β (log fee) Std. Err. t-stat p-value R2

Base (Flashblocks) 1 -0.005 0.001 -9.424 < 0.001 0.791
2 -0.001 0.000 -3.558 < 0.001 0.186
3 0.000 0.000 0.938 0.353 0.028
4 -0.000 0.000 -1.272 0.210 0.064
5 0.000 0.000 0.627 0.534 0.017
6 -0.000 0.000 -0.904 0.371 0.035
7 -0.000 0.000 -0.602 0.550 0.015
8 0.000 0.000 0.856 0.396 0.028
9 0.000 0.000 1.777 0.081 0.100
10 -0.002 0.000 -5.672 < 0.001 0.434
Unichain (Flashblocks) 1 -0.001 0.000 -4.163 < 0.001 0.366
2 -0.001 0.000 -1.986 0.052 0.114
3 0.000 0.000 0.379 0.706 0.005
4 -0.000 0.000 -0.144 0.886 0.000
5 0.000 0.000 0.521 0.605 0.008
6 -0.001 0.000 -3.050 0.004 0.206
7 -0.001 0.000 -3.157 0.003 0.217
8 -0.000 0.000 -0.393 0.696 0.005
9 -0.001 0.000 -4.507 < 0.001 0.418
10 -0.000 0.000 -1.931 0.060 0.110
Table 3: Slot-level regressions of median transaction position on log(priority fee)
for Base and Unichain under Flashblocks. Strong fee-ordering persists in some
slots (e.g., Base slots 1 and 10; Unichain slots 1, 6, 7, 9), while others show weak
or no significant relationship. This reflects how Flashblocks localizes priority fee
auctions to specific slots, reducing global fee-ordering efficiency.

Chain β (log fee) Std. Err. t-stat p-value R2

Arbitrum (pre-TimeBoost) -0.001 0.002 -0.367 0.715 0.004
Arbitrum (TimeBoost) -0.000 0.002 -0.153 0.879 0.001
Base (pre-Flashblocks) -0.058 0.003 -22.053 < 0.001 0.853
Unichain (pre-Flashblocks) -0.047 0.002 -23.219 < 0.001 0.919
Optimism -0.053 0.002 -25.961 < 0.001 0.882
ZKsync -0.004 0.006 -0.618 0.540 0.014
Table 4: Pooled OLS regressions of median transaction position on
log(priority fee) across L2s. Negative coefficients indicate that higher fees lead
to earlier block positions. Results show efficient fee-based ordering on Base,
Unichain, and Optimism, while Arbitrum and ZKsync exhibit no significant re-
lationship. This is consistent with their FCFS (first-come, first-served) sequencer
policy rather than a priority fee auction (PFA).
 Revert-Based MEV on Fast-Finality Rollups 25

B Proofs

B.1 Proof of Proposition 1

To prove this proposition, we first need two lemmas.

Lemma 1. Under same-pool, same-block execution with no external state re-

fresh, the total output of n equal slices sums to the output of one big trade:
n
X
∆yi = ∆y(D).
i=1

Proof. Consider a constant-product AMM with reserves (x, y) and invariant k =

k
xy. A single trade’s payout is ∆y(q) = y − x+(1−f )q , consistent with (1). A single
k (1−f )D
trade of (pre-fee) size D therefore yields ∆y(D) = y − x+(1−f )D . Let d := n
denote the post-fee input per slice. After i − 1 slices the x-reserve is x + (i − 1)d
k k
and the y-reserve is k/(x+(i−1)d). The i-th slice outputs ∆yi = x+(i−1)d − x+id .
Summing telescopes,
n      
X k k k k k
∆yi = y− + − + ··· + −
i=1
x+d x + d x + 2d x + (n − 1)d x + nd
k k
=y− =y− = ∆y(D).
x + nd x + (1 − f )D

⊔
⊓

Remark 3. Lemma 1 assumes the n slices are executed atomically, so the pool
follows one deterministic path. This is a standard simplification to isolate the
frictionless AMM mechanics. Any sandwiching or visibility costs are modeled
separately by the convex loss term, not inside the mechanical output.
If interleaving is possible (public mempool), the proof still goes through with
a small tweak: decompose the outcome of splitting into (i) a path effect and (ii)
an adversarial-loss effect.

Lemma 2. If L is strictly convex with L(0) = 0, then for any D > 0 and integer
n > 1,
L(D) − n L(D/n) > 0.

Proof. Let λ = 1/n
∈ (0, 1) and x = D. By strict convexity of L and L(0) = 0,

L λx + (1 − λ) · 0 < λL(x) + (1 − λ)L(0) = n1 L(D). Since λx + (1 − λ) · 0 = λD =
D/n, this is L(D/n) < n1 L(D). Multiplying by n yields n L(D/n) < L(D), i.e.,
L(D) − nL(D/n) > 0. ⊓
⊔

Now we can proceed to prove the proposition.

26 K. Gogol, M. Schneider, C. Tessone

Proof of Proposition 1 Compute the difference directly:

" n #
X


Π(n) − Π(1) = s∆yi − Pc D/n − L(D/n) − κ − s∆y(D) − Pc D − L(D) − κ
i=1
n
!
X

=s ∆yi − ∆y(D) + L(D) − nL(D/n) − (n − 1)κ,
i=1
Pn
which is (2). Under same-pool, same-block execution, Lemma 1 yields i=1 ∆yi =
∆y(D), so the first bracket vanishes and we obtain


Π(n) − Π(1) = L(D) − nL(D/n) − (n − 1)κ,

as in (3). By Lemma 2, L(D) − nL(D/n) > 0 for n > 1, hence Π(n) > Π(1)
whenever L(D) − nL(D/n) > (n − 1)κ. ⊓
⊔

B.2 Proof of Proposition 2

Proof. Write g(K) := 1−(1−s)K . Then FL (K; q) = g(K) A−Kcg . The discrete
marginal gain from adding one more copy is

∆K FL = FL (K; q)−FL (K −1; q) = g(K)−g(K −1) A−cg = (1−s)K−1 sA−cg .

(i) If A ≤ 0, then FL (K; q) ≤ −Kcg and K ⋆ = 1. If A > 0 but sA ≤ cg , then

∆1 FL = sA − cg ≤ 0 and the objective decreases for all K ≥ 2, hence K ⋆ = 1.
(ii) If A > 0 and sA > cg , then ∆1 FL > 0 and ∆K FL is strictly decreasing
in K because (1 − s)K−1 decreases in K. Thus FL is strictly increasing until the
smallest K for which ∆K FL < 0, and strictly decreasing thereafter. The optimal
K ⋆ is the largest integer with ∆K FL ≥ 0, which rearranges to (5). ⊔
⊓
 Revert-Based MEV on Fast-Finality Rollups 27

C Additional Figures

Arbitrum Base Optimism Unichain ZKsync

100%
92.19
80%
Priority Fee Per Gas

73.42
60%
58.14
54.23
40% 47.34 47.9
35.8
20% 24.36
19.36 21.31
4.58 5.33 2.22
9.79 0.08 0.09 1.2 0.18 0.16 2.07 0.25
0%
null

0 wei

1 wei

<=1 gwei

>1 gwei

null

1 wei

<=1 gwei

>1 gwei

null

1 wei

<=1 gwei

>1 gwei

null

0 wei

1 wei

<=1 gwei

>1 gwei

null

0 wei

<=1 gwei

>1 gwei
Fig. 8: Priority fee per gas in reverted transactions across L2s. Base, Optimism,
and Unichain implement PFAs, whereas Arbitrum and ZKsync rely on FCFS
ordering. Strikingly, nearly half of reverted transactions—and almost all on
ZKsync—include a priority fee despite the absence of a priority fee auction at
these L2s.
28 K. Gogol, M. Schneider, C. Tessone

Gas Used Total Fee

Arbitrum
% from Reverted Transactions

60

40

20

0
Jan 2025 Feb 2025 Mar 2025 Apr 2025 May 2025 Jun 2025 Jul 2025

Base
% from Reverted Transactions

30

20

10

0
Jan 2025 Feb 2025 Mar 2025 Apr 2025 May 2025 Jun 2025 Jul 2025

Optimism
80
% from Reverted Transactions

60

40

20

Jan 2025 Feb 2025 Mar 2025 Apr 2025 May 2025 Jun 2025 Jul 2025

Unichain
60
% from Reverted Transactions

50

40

30

20

10

Jan 2025 Feb 2025 Mar 2025 Apr 2025 May 2025 Jun 2025 Jul 2025

ZKsync
% from Reverted Transactions

25

20

15

10

0
Jan 2025 Feb 2025 Mar 2025 Apr 2025 May 2025 Jun 2025 Jul 2025

Fig. 9: Daily share of reverted transactions in total gas consumption and fee
revenue on each L2. Reverted transactions contribute disproportionately more
to L2 fee revenues than to gas usage
 Revert-Based MEV on Fast-Finality Rollups 29

Priority Fee Base Fee L1 Fee

Arbitrum
400
Total Daily Gas Fee [ETH]

300

200

100

0
Jan 12 Jan 26 Feb 9 Feb 23 Mar 9 Mar 23 Apr 6 Apr 20 May 4 May 18
2025

Base
Total Daily Gas Fee [ETH]

300

200

100

0
Jan 12 Jan 26 Feb 9 Feb 23 Mar 9 Mar 23 Apr 6 Apr 20 May 4 May 18
2025

Optimism
Total Daily Gas Fee [ETH]

40

30

20

10

0
Jan 12 Jan 26 Feb 9 Feb 23 Mar 9 Mar 23 Apr 6 Apr 20 May 4 May 18
2025

Unichain
Total Daily Gas Fee [ETH]

10

0
Jan 12 Jan 26 Feb 9 Feb 23 Mar 9 Mar 23 Apr 6 Apr 20 May 4 May 18
2025

ZKsync
Total Daily Gas Fee [ETH]

15

10

0
Jan 12 Jan 26 Feb 9 Feb 23 Mar 9 Mar 23 Apr 6 Apr 20 May 4 May 18
2025

Fig. 10: Daily total gas fees incurred by reverted transactions. The breakdown
shows the L2 execution fee (split into base and priority components) and the L1
fee for calldata posting to Ethereum.
30 K. Gogol, M. Schneider, C. Tessone

Arbitrum (pre TimeBoost) Base (pre Flashblocks) Unichain (pre Flashblocks)

61 1.0 61 1.0 61 1.0

51 52 52
0.8 0.8 0.8
Priority Fee per Gas (wei)

42 43 44
0.6 0.6 36 0.6
34 34

25 0.4 26 0.4 27 0.4

17 17 19
0.2 0.2 0.2
8 8 11

0 0.0 0 0.0 0 0.0

1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Arbitrum (TimeBoost) Base (Flashblocks) Unichain (Flashblocks)
61 0.8 61 1.0 61 1.0

50 0.7 52 52
0.8 0.8
0.6
Priority Fee per Gas (wei)

42 43 43
0.5 0.6 0.6
33 34 34
0.4
25 26 0.4 26 0.4
0.3
16 17 17
0.2
0.2 0.2
8 0.1 8 8
0 0.0 0 0.0 0 0.0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Arbitrum (pre TimeBoost) Base (pre Flashblocks) Unichain (pre Flashblocks)
Zero-fee tx (bin 0) 25 Zero-fee tx (bin 0) 50 Zero-fee tx (bin 0)
30 Positive-fee tx (bins 1) Positive-fee tx (bins 1) Positive-fee tx (bins 1)

25 20 40
Revert Share (%)

20 15 30

15
20
10
10
10
5 5
0
2 4 6 8 10 2 4 6 8 10 2 4 6 8 10
Arbitrum (TimeBoost) Base (Flashblocks) Unichain (Flashblocks)
Zero-fee tx (bin 0)
Positive-fee tx (bins 1) 14
25 16
12

14 10
Revert Share (%)

20
8
15 12
6
4
10 10
2
Zero-fee tx (bin 0) Zero-fee tx (bin 0)
5 8 Positive-fee tx (bins 1) 0 Positive-fee tx (bins 1)
2 4 6 8 10 2 4 6 8 10 2 4 6 8 10
Block Slot of Tx Index Block Slot of Tx Index Block Slot of Tx Index

Fig. 11: Heatmaps of revert share (y-axis, color intensity) across log priority fee
levels (x-axis) and Flashblock/TimeBoost slots (vertical axis), with slot-level
median position vs. fee lines overlaid. Panels show before vs. after sequenc-
ing changes on Base, Unichain (Flashblocks), and Arbitrum (TimeBoost). The
heatmaps reveal localized auction behavior at Slot 1 after the mechanisms are
introduced, while Slots 2–9 remain insensitive to fees. Overlaid line plots high-
light slot-by-slot fee responsiveness, confirming that only the block head exhibits
auction-like dynamics.