# RANDAO Manipulation in the Presence of MEV

Crypto.world

> Ethereum selects each epoch's block proposers using RANDAO, a randomness beacon that a proposer can influence by choosing whether to propose or skip their assigned slots — particularly the…

**arXiv**: 2610.02143

**Published**: 2026-10-01

**Authors**: Kaya Alpturer, Nicholas Hope, S. Matthew Weinberg

## AI summary

## Overview

**Research area:** Algorithmic game theory and mechanism design for blockchain protocols — specifically, the strategic manipulation of Ethereum's on-chain randomness beacon (RANDAO) when block proposers earn variable execution-layer rewards (MEV) rather than a flat per-block reward.

**Technical level:** Advanced. The paper builds on Markov Decision Process (MDP) formulations of validator strategy, proves truncation and concentration bounds, and uses a sampling-based approximate solver. The motivation and qualitative conclusions are accessible to a non-specialist, but the methodology and guarantees require comfort with stochastic processes and game theory.

**Scope (one sentence):** The paper presents a sampling-based method for computing ε-optimal RANDAO-manipulation strategies under arbitrary reward models — including reward rollover, multi-block MEV, CEX-DEX arbitrage, and oracle manipulation — and applies it to quantify how reward structure changes the incentive to manipulate, plus how epoch length and tail slashing can mitigate it.

## What This Paper Is About

Ethereum selects each epoch's block proposers using RANDAO, a randomness beacon that a proposer can influence by choosing whether to propose or skip their assigned slots — particularly the consecutive slots at the end of an epoch (the "tail"), where an adversary controlling *t* slots can select among 2^*t* possible futures. Prior work computed the fraction of blocks a strategic proposer can capture, but that fraction is only a proxy for what validators actually care about: rewards.

This paper asks how the incentive to manipulate RANDAO changes when rewards are not uniform — when they roll over across missed slots, or grow super-linearly with runs of consecutive blocks, as with arbitrage, sandwich attacks, TWAP-oracle manipulation, and priority fees. The goal is a general method that computes an ε-optimal manipulation strategy for essentially any reward scheme expressible as a function of the adversary's proposal pattern, and then to use it to characterize when manipulation becomes more or less attractive.

## Key Contributions

1. **A general reward model for RANDAO manipulation.** The authors define a reward model as a pair of functions, `propose(·)` (expected reward from a fully specified tail proposal pattern, where unproposed tail slots are *missed*) and `next(·)` (expected reward from a non-tail proposal pattern, where unproposed slots are proposed *honestly*). They show this pair captures consensus-reward rollover, per-slot MEV with rollover, smooth multi-block MEV (arbitrage-like), threshold multi-block MEV (oracle-manipulation-like), and any combination of these.
2. **A provable truncation of the RANDAO game plus a sampling-based solver.** Because transition probabilities become combinatorially intractable once rewards depend on the *pattern* rather than the *count* of proposals, the authors truncate the state space to tails {0, …, *T*} and estimate transitions empirically. Upper and lower truncations bracket the true optimum, giving rigorous bounds on the approximation.
3. **Robustness guarantees.** They prove a bound on the expected number of consecutive rounds spent in large tails (Lemma 5), an optimality sandwich theorem for the truncated MDPs (Theorem 6), and a Chernoff-Hoeffding-based sample-complexity bound (Lemma 7) for estimating transitions and rewards to within ε with probability 1 − δ. For ε = 0.0085 and δ = 0.01, N = 10⁸ samples per state suffices for T = 7.
4. **Applied analyses and a public codebase.** They reproduce prior uniform-reward results, run three execution-layer reward models, vary epoch length, and encode tail-slashing penalties as a reward function to find penalties under which honest behavior is ε-optimal. Code is released at https://github.com/kalpturer/randao-mev.

## Main Findings

- **The methodology reproduces the known uniform-reward benchmark.** For a single unit of reward per proposed block, the solver's upper (*U*) and lower (*L*) truncation bounds bracket the values reported in prior work [2], with each entry carrying an error of at most 0.0265% (ε/32): | α | 5% | 10% | 15% | 20% | 25% | 30% | 35% | | [2] | 5.0483% | 10.1881% | 15.3996% | 20.6777% | 26.0247% | 31.4516% | 36.9734% | | U | 5.0483% | 10.1879% | 15.3998% | 20.6783% | 26.0323% | 31.4916% | 37.2290% | | L | 5.0483% | 10.1878% | 15.3997% | 20.6776% | 26.0250% | 31.4487% | 36.9581% |
- **Partial reward rollover amplifies the manipulation incentive.** In the per-slot MEV model (1 unit of reward per slot, split 60% ephemeral / 40% rollover, with R_max = 32), the strategic proposer's reward exceeds the honest baseline at every stake level tested. At α = 35%, the upper bound is 37.3514% and the lower bound 37.0754%, versus 35.0016% for the honest strategy; at α = 5%, upper and lower are both 5.0493% against honest 4.9999%.
- **Super-linear scaling with consecutive blocks amplifies the incentive further.** Two stylized multi-block models are used: a *smooth* model paying *k*√*k* for *k* consecutive blocks (motivated by CEX-DEX arbitrage, where fees create a no-trade band and an option-like payoff), and a *threshold* model paying *k* for *k* < 5 and *k* + 10 for *k* ≥ 5 (motivated by TWAP-oracle manipulation enabling forced liquidations). Both are explicitly described as stylized illustrations rather than calibrated estimates; the specific quantitative results for these models fall in the truncated portion of the provided content and are not reported here.
- **Effect size depends heavily on the reward model.** The headline qualitative conclusion is that "the effect of RANDAO manipulation is highly dependent on the underlying reward model" — the block-fraction proxy used in prior work understates the payoff when rewards roll over or scale with block runs.
- **Longer epochs reduce manipulation gains across reward models.** Prior work [2] observed this for uniform rewards; the authors extend the finding to non-uniform reward structures, strengthening the case for longer epochs as a mitigation.
- **Tail slashing can restore honest equilibria.** Encoding a slashing curve that penalizes misses toward the tail more heavily and amplifies penalties for groups of consecutive misses as a reward function, the authors compute — for each MEV reward model — minimal slashing curves under which honest behavior is ε-optimal.
- **Empirical caveats are stated explicitly.** Consensus-reward rollover is directly visible in data, but MEV rollover is harder to verify because slot misses are rare on Ethereum, and systematic multi-block MEV is described as weak in current Ethereum [26]. The authors note the modelled MEV regimes may matter more under future block-building redesigns.

## Methodology in Plain English

**The game.** Ethereum's timeline is divided into 32-slot epochs (12 seconds per slot). At the end of each epoch, the accumulated RANDAO value determines who proposes in the next epoch. A validator who controls the last *t* slots of an epoch knows, in advance and privately, the exact randomness each of those slots would contribute — so by choosing which subset to actually propose, they can pick their favourite among 2^*t* possible futures for the next epoch's 32 proposers. The price of this choice is whatever rewards they forgo by skipping.

**Modeling rewards.** The authors note that for a strategic proposer, only two things matter: the reward from the tail pattern they choose, and the reward from the non-tail pattern of an epoch they did not control. Any reward scheme expressible this way can be plugged in. This is a deliberate departure from prior work, which needed uniform rewards to collapse the analysis.

**Why the old approach breaks.** Computing transition probabilities between tail lengths requires summing over every possible reward realization the adversary might face. With uniform rewards, the optimal action depends only on *how many* slots you propose; with pattern-dependent rewards, it depends on *which* slots, and the sum explodes (the smooth model, for instance, requires enumerating integer partitions of every proposal count). Sampling sidesteps this entirely.

**The solver.** States are tail lengths {0, …, 32}. The method (1) truncates to {0, …, *T*}, redirecting any transition past *T* back to *T*, with the reward of that transition set to either an upper bound (R_max · E[L(T, α)]) or zero; (2) estimates the truncated transition matrix and reward vector from N samples per state; (3) computes the average reward via the stationary distribution of the estimated chain; and (4) performs policy improvement by solving V + γ = R̂ + P̂V, iterating until values stop moving.

**Why the bounds hold.** The key structural fact is that any excursion into very large tails must eventually return to a small tail, and the policy that always picks the longest tail stochastically dominates every other policy in time spent in large tails. That bounds the total reward attributable to the discarded states, and the two truncations (upper and lower) sandwich the true optimum. Sampling error is controlled by a Chernoff-Hoeffding union bound over the (T+1)² transition entries and (T+1) reward entries.

**Experimental setup.** Experiments focus on α ≤ 0.35 because beyond that the expected time spent in large tails grows rapidly and the truncation bound separating upper and lower estimates degrades. Assumptions include i.i.d. uniform proposer sampling, honest validators always proposing, the adversary always proposing in non-tail slots, reward "resets" after an honest proposer, and ignoring the real-world two-epoch interleaving (RANDAO from epoch *i* selects proposers for epoch *i*+2).

## Why This Matters

**Impact on research.** The paper moves the RANDAO-manipulation literature from "how many blocks can an adversary capture" to "how much can an adversary earn," which is the quantity validators actually optimize. It supplies a reusable computational toolkit for arbitrary reward functions, so new MEV structures can be analyzed without re-deriving closed forms, and it extends a known finding from the selfish-mining literature — that non-uniform fees reshape equilibria [10] — to proof-of-stake randomness manipulation.

**Real-world applications (grounded in the paper's examples):**

- **Validator and staking-pool economics:** with Ethereum's stake pool at roughly $80 billion, small shifts in achievable reward fractions translate into large absolute sums, and operators deciding whether to run honest or optimizing infrastructure need to know the gap.
- **CEX-DEX arbitrage:** an adversary holding a run of consecutive blocks can execute a single larger arbitrage more profitably than several small ones, because DEX fees (e.g., the paper's 30 basis point example) create a profitability threshold — making long runs disproportionately valuable.
- **TWAP-oracle manipulation:** controlling consecutive blocks lets an attacker push a time-weighted average price far enough to trigger illegitimate liquidations, with RANDAO manipulation lowering the cost by giving control over transaction inclusion.
- **Protocol parameter design:** epoch length and missed-slot / tail-slashing penalties are both live design levers, and the paper quantifies how each shifts the manipulation incentive.

**Industry relevance.** The results bear directly on decisions made by client teams and protocol researchers: whether to lengthen epochs, whether to introduce targeted missed-slot penalties (a topic of community discussion [22]), and how to reason about proposer-builderseparation and block-building redesigns that could make multi-block MEV regimes more relevant than they are today.

## Future Directions

- **Empirically calibrating the rollover model.** The authors note that MEV rollover is hard to verify from chain data because slot misses are rare, and that systematic multi-block MEV is weak in current Ethereum [26]. Measuring what actually happens to MEV opportunities across a missed slot would replace the stylized parameters with observed ones.
- **Stress-testing under block-building redesigns.** The smooth and threshold multi-block models are explicitly stylized; the paper argues these regimes "may matter in the future under block-building redesigns," which invites re-running the methodology under proposed architectures that concentrate consecutive block control.
- **Designing implementable slashing curves.** The paper computes, per MEV reward model, minimal slashing curves that make honesty ε-optimal. Translating these into penalties that are incentive-compatible under real-world conditions — and cheap enough not to punish benign outages — is a natural next step.
- **Relaxing the modeling assumptions.** The analysis assumes i.i.d. proposer selection, no epoch interleaving, honest proposers always proposing, and reward independence across honest proposals. Each relaxation (particularly adaptive strategies that condition on realized fee draws, as in [6, 10]) is a substantive open problem.

*Note: the supplied paper content is truncated partway through Section 4.3, so the authors' own Section 7 future-work list and the specific numeric results for the multi-block and slashing experiments are not available here and are not reported above.*

## Target Audience

Protocol researchers and mechanism designers working on proof-of-stake consensus and MEV; algorithm-game-theory researchers interested in MDP-based analysis of blockchain deviations; validator operators and staking-pool engineers who want to understand the magnitude of the manipulation gap; and graduate students with a background in stochastic processes or game theory who are entering the blockchain-economics literature. Readers seeking an intuitive, non-technical account will find Sections 1 and 4.1–4.2 accessible; the guarantees in Section 3 require more mathematical maturity.

## Authors’ abstract

Ethereum's randomness beacon (RANDAO) is well-known to be manipulable, and prior work [AW24] computes the precise fraction of blocks a strategic proposer can propose. The fraction of blocks proposed, however, is only a proxy for participants' rewards. We propose a generalized reward model capturing many canonical forms of rewards: consensus reward rollover, multi-block MEV, CEX-DEX arbitrage, oracle manipulation, and others. We provide a methodology that computes an…

[The full abstract on arXiv](https://arxiv.org/abs/2610.02143)

[Read the original paper](https://arxiv.org/abs/2610.02143)

---

Canonical: https://crypto.world/paper/randao-manipulation-in-the-presence-of-mev-2610.02143
