Research
Imperfect Commitment in Maximal Extractable Value Auctions
Ethereum block builders run sealed auctions in which searchers submit bundles with hidden tips, and the builder picks winners. Nothing in the protocol forces the builder to honor that outcome once it…

- arXiv
- 2605.22667
- Published
- 2026-05-21
- Authors
- Aleksei Adadurov, Sergey Barseghyan, Anton Chtepine, Antero Eloranta, Andrei Sebyakin, Arsenii Valitov
AI summary
Overview
Research area: Blockchain economics / mechanism design — specifically Maximal Extractable Value (MEV) auctions on Ethereum, sitting at the intersection of auction theory, game theory, and empirical crypto-market analysis.
Technical level: Intermediate. The empirical sections are accessible to anyone familiar with Ethereum block-building, but the equilibrium analysis assumes comfort with Bayesian Nash equilibrium, affiliated values, and first-price auction theory.
Scope: The paper models and empirically measures what happens when an Ethereum block builder can secretly front-run the winning searcher in its own sealed-bid MEV auction, using 2.2 million MEV bundle transactions to estimate how much surplus is exposed to this behavior across four MEV types.
What This Paper Is About
Ethereum block builders run sealed auctions in which searchers submit bundles with hidden tips, and the builder picks winners. Nothing in the protocol forces the builder to honor that outcome once it has seen every bundle's contents. The paper asks how much this commitment problem costs: it models a builder that defects with some probability, estimates how much of each MEV opportunity a defecting builder could actually replicate, and measures the gap between what searchers currently pay and what a defecting builder could take.
Key Contributions
-
Isolates the builder-defection channel from the auction-format question. The authors position their work explicitly as a companion to prior work on which auction format an honest builder should choose, arguing that even a format-optimal sealed auction is undermined when the builder can act on observed bid information after the fact.
-
Characterizes a piecewise symmetric Bayesian Nash equilibrium under partial commitment. Searchers with valuations below a type-specific cutoff bid as in a standard affiliated first-price auction; above the cutoff they jump to a deterrence bid equal to the replicable fraction of their value, making front-running unprofitable for the builder.
-
Shows the builder's optimal defection rate has only boundary solutions. Under a high-extractability regime the builder's expected revenue is strictly increasing in the defection rate, so the optimum is full defection; under a low-extractability regime the optimum is zero defection. There is no interior optimum in the baseline model.
-
Estimates the type-specific replicability parameter from the data and decomposes revenue. Using right-tail plateaus in bribe schedules across four MEV types, the paper recovers type-specific replicability estimates and compares observed tips against the surplus a defecting builder could capture, finding sharp heterogeneity.
Main Findings
-
Commitment cost is highly type-specific. The estimated foregone front-run surplus across all types is $49.4M, or 48.8% of observed tips. The burden is not evenly distributed: naked arbitrage contributes $24.3M of that against $18.3M of observed tips, and liquidations contribute $12.4M against $7.0M of observed tips.
-
Sandwiches are already saturated by competition. Sandwich bribe shares sit at roughly 90–95% across the value range with a mean bribe share of 95%, so there is no sharp structural break to detect and the front-running constraint rarely binds. Estimated foregone surplus is only $7.3M on $57.8M of observed tips.
-
Naked arbitrage shows the clearest transition. The bribe share rises from roughly 45% at low values to about 75% in the right tail, giving an estimate of γ̂ = 0.74, which exceeds the mean bribe share of 67% and leaves a gap of about 30 percentage points above the observed bid share in the low-value region.
-
Liquidations are thin, large, and exposed. The bribe schedule rises from roughly 55% at low values to 85–90% in the right tail, with γ̂ = 0.88 against a mean bribe share of 68%. Average foregone surplus per liquidation is approximately $2,600, on only 4,759 observations.
-
Backruns are intermediate. The bribe schedule is flat at roughly 70% with a slight decline in the highest-value bins — a decline the baseline model does not predict. Foregone surplus is $5.5M on $18.3M of observed tips.
-
The risky-regime bid function is independent of the defection probability. Adding the defection probability scales a searcher's expected payoff uniformly, so it does not change the first-order condition; only the choice between the risky and safe regimes depends on it.
-
The honest-disclosure benchmark is an upper bound, not a target. The revenue-maximizing disclosure threshold for an honest auctioneer depends on the effective number of bidders, which is type-specific. The authors argue that once defection is possible, admissible information exposure must sit below that honest benchmark wherever the defection gain is positive.
-
Affiliation in the data is consistent with the model's assumption. Within each block and MEV type, the log extracted value of the largest opportunity is positively related to that of the second-largest, supporting the maintained assumption that searcher valuations are not independent draws.
Methodology in Plain English
The model. A single builder runs a sealed first-price auction among risk-neutral searchers who compete for one MEV opportunity. Each searcher gets a noisy private signal about the opportunity's value, and values are correlated across searchers (captured by an affiliation parameter, with common information such as victim transaction size or price discrepancies). The builder defects with probability ε. On defection, it uses the observed bundle payloads to replicate the winner's strategy and capture a fraction γ(τ) of the opportunity, where τ is the MEV type. Standard revert protection means the searcher's bundle fails and no payment is collected, so the builder's decision is simple: front-run if and only if γ(τ)v exceeds the winning bid.
The equilibrium. A searcher can either bid low and risk being front-run, or bid the deterrence level γ(τ)v, which removes the builder's incentive to defect. Below a cutoff valuation, the standard affiliated first-price bid is optimal; above it, the deterrence bid wins out. The cutoff is where the two expected payoffs are equal. The authors note that the cutoff truncates the distribution of rival bids, but argue this effect is second-order because the cutoff lies deep in the right tail of a log-normal with σ̂ ≈ 2.524.
The data work. The paper uses the libmev dataset: 2.2 million MEV bundle transactions on Ethereum from September 2024 to August 2025, with $168.5M in total extracted value and $101.3M in tips to builders. Extracted value is defined as tip plus retained profit, so the observed bribe share is the empirical counterpart of the model's bid-to-value ratio. Each panel plots average bribe share against log extracted value using 50 quantile bins; the right-tail plateau in the top 20% of bins is read off as the estimate of γ(τ). The model's log-normal fit uses μ̂ = 1.102 and σ̂ = 2.524. The builder field identifies five major builders accounting for 93% of transactions: beaverbuild.org, Titan, BuilderNet (Beaver), bobTheBuilder, and BuilderNet (Flashbots).
The decomposition. For each transaction, potential defection gain is the positive part of γ̂(τ)v minus the observed bribe, aggregated by MEV type. This is the data analogue of the difference between the defection and no-defection terms in the builder's expected revenue.
Why This Matters
For research: The paper opens a channel that auction-format research treats as external. It argues that the design question is not "which auction format" but "how do you constrain what the builder can do with information it has already seen" — and it provides a parameter (γ(τ)) that makes the answer measurable rather than rhetorical. It also connects the classic corrupt-auctioneer literature, which is silent on which manipulation channel dominates in any specific institutional setting, to a concrete market where the answer can be estimated.
Real-world applications:
- Relay and PBS design: Safeguards that delay, aggregate, or withhold the winning payload and bidder identity at the moment of allocation, rather than changing the auction format.
- Builder accountability: Since defection is only costly if observable, the framework points toward mechanisms that make front-running detectable and punishable rather than merely discouraged.
- Searcher bidding strategy: The piecewise equilibrium gives searchers a concrete rule for when to shade and when to bid up to the deterrence threshold, differentiated by MEV type.
- Protocol-level MEV policy: The finding that liquidations and naked arbitrage carry far more exposed surplus than sandwiches suggests where intervention would matter most.
Industry relevance: The dataset is dominated by five builders responsible for 93% of transactions, and the authors note that builder centralization concentrates discretion in a small number of agents. Orderflow access and integrated searcher-builder structures create persistent informational asymmetries, which sharpens the corruption analogy for anyone operating in the block-building layer.
Future Directions
-
Estimate ε itself. The paper treats the defection probability as an exogenous builder property and does not report an empirical estimate of it; the authors explicitly flag estimating ε and γ(τ) at the builder-searcher level as future work.
-
Model selective disclosure to integrated searchers. When a searcher is integrated with the builder, defection becomes selective disclosure to a favored bidder rather than outright replication. The authors mention this as a natural extension but do not analyze it in the baseline model.
-
Explain the backrun right-tail decline. Backrun bribe shares fall slightly in the highest-value bins, which the baseline model does not predict; the authors suggest this may reflect specialization among high-value opportunities, leaving it as an open empirical question.
-
Test commitment-preserving mechanisms. The paper calls for testing mechanisms that preserve competition while restricting ex-post use of bid and payload information.
Target Audience
Researchers in mechanism design and game theory working on blockchain markets; Ethereum protocol and PBS infrastructure designers; builder and relay operators who need to reason about credibility of their own auctions; searchers and MEV analysts interpreting bribe schedules across opportunity types; and policy-oriented readers interested in where MEV market failures are concentrated. Readers without auction-theory background can extract most of the value from the data section and findings, but the equilibrium characterization in Section III requires familiarity with affiliated-values first-price auctions.
Authors’ abstract
Ethereum block builders run sealed auctions among searchers, but nothing in the protocol forces a builder to honor the auction outcome after observing submitted bundles. This paper studies the commitment problem. We model a builder who defects with probability $\varepsilon$ and, upon defection, replicates a type-specific fraction $γ(τ)$ of the winning MEV opportunity. Searchers anticipate this behavior and choose between a risky first-price bid and a safe deterrence bid that…
The full abstract on arXiv