Are AMMs Being Designed for the Wrong Primary User?

Magicians
,
1

The implicit assumption behind every major AMM design is a mixed flow: retail users swapping tokens for actual use, and arbitrageurs correcting price discrepancies. This assumption was never stated explicitly, but it shaped every architectural decision — pair pools, flat fees, uniform liquidity ranges, anti-MEV mechanisms.

I’ve been questioning whether that assumption still holds as the primary design constraint.

The empirical picture is harder to read than it first appears. Heimbach et al. [1] attribute over 25% of volume on Ethereum’s five largest DEXes to non-atomic arbitrage alone. Canidio and Fritsch [2] estimate combined arbitrage at roughly 40% of Uniswap v3 volume, leaving ~60% classified as “noise traders.” Qin et al. [3] quantified hundreds of millions in extracted value from bot-originated activity across DEXes. On the surface this suggests organic retail flow is still the majority.

But I think there’s a methodological problem with how “real user” gets defined in this literature. These papers identify bots by looking for known arbitrage patterns — high frequency, cross-venue price correction, atomic execution. Everything that doesn’t match those patterns gets classified as organic flow by residual. There’s no independent verification that the residual actually represents human-initiated activity.

Consider what that residual probably contains: bots executing swaps on behalf of users through aggregators and intent systems (on-chain the transaction is bot-originated, but it gets counted as organic); low-frequency bots that don’t exhibit the high-frequency signatures used to identify arbitrageurs; protocol rebalancing bots; and noise-trading bots that deliberately mimic retail patterns. On-chain data has no reliable signal for human intent — every transaction looks identical regardless of whether a human clicked a button or a script fired autonomously.

There’s also a more fundamental categorization problem: aggregators and arbitrageurs are functionally identical from the AMM’s perspective. Both route flow toward the best available price. Both extract value from a pool when it’s mispriced relative to other venues. The AMM receives the same transaction either way and cannot distinguish between them. The difference is meaningful from the end user’s perspective — one is executing a human’s intent, the other is capturing a price discrepancy for its own account — but it’s invisible to the AMM’s fee and pricing logic. Classifying aggregator-routed volume as “organic” while classifying direct arbitrage as “bot activity” creates an artificial distinction that doesn’t reflect how either actor interacts with the pool.

The deeper observation is simpler: humans are not very active on-chain directly. The friction of gas costs, wallet management, and on-chain UX creates a high bar for unmediated activity. Most humans who interact with DeFi do so through an interface that abstracts the execution — which means the actual on-chain actor is almost always some form of automated system, whether that’s an aggregator solver, an intent filler, or a classic arbitrage bot.

This creates a tension I keep coming back to:

  • A large and likely undercounted fraction of on-chain AMM flow is automated

  • The dominant research direction is making life harder for automated actors

  • LPs are evaluated against benchmarks like LVR [4] and IL that measure relative performance against an idealized market, not whether their USD position actually grew

  • The “organic flow” that supposedly subsidizes LPs is increasingly intermediated by aggregators and batch systems before it reaches the AMM directly

And yet the design conversation keeps centering retail protection as the primary objective, with arbitrageurs as the adversary to mitigate.

I’m not sure this framing is entirely wrong — retail users do deserve good execution when they show up. But I wonder if the assumption that organic flow is the majority is less well-supported than the literature suggests, and whether that changes the design priorities.

If automated flow is larger than we think, what does that change about how you’d design fee mechanisms, liquidity structure, or LP incentives?

And separately: is LVR the right benchmark for LP profitability, or is it answering a different question than what a real LP actually cares about? Milionis et al. [4] are explicit that LVR measures costs relative to a rebalancing strategy, not absolute USD return — which raises the question of whether it’s the right objective function for a passive LP with a finite horizon.

Curious if others have better methodologies for distinguishing human from automated flow, or data that pushes back on this framing.

References

[1] L. Heimbach, V. Pahari, and E. Schertenleib, “Non-Atomic Arbitrage in Decentralized Finance,” in IEEE Symposium on Security and Privacy (SP), San Francisco, CA, USA, May 2024. arXiv:2401.01622.

[2] A. Canidio and R. Fritsch, “Arbitrageurs’ Profits, LVR, and Sandwich Attacks: Batch Trading as an AMM Design Response,” arXiv preprint arXiv:2307.02074, 2023.

[3] K. Qin, L. Zhou, and A. Gervais, “Quantifying Blockchain Extractable Value: How Dark is the Forest?” in 2022 IEEE Symposium on Security and Privacy (SP), IEEE, 2022, pp. 198–214. DOI: 10.1109/SP46214.2022.9833734.

[4] J. Milionis, C. C. Moallemi, T. Roughgarden, and A. L. Zhang, “Automated Market Making and Loss-Versus-Rebalancing,” arXiv preprint arXiv:2208.06046, 2022.

Tags: amm-design, mev, fee-mechanism, liquidity-providers, defi-research

1 Like
2

Re: Are AMMs Being Designed for the Wrong Primary User?

Really appreciate the methodological sharpness here. The residual-classification critique is underappreciated — it’s easy to mistake “not-arbitrage” for “retail” without ever establishing that independently.

A few questions and observations, offered in the spirit of pushing the argument further:

If the distinction is unmeasurable, what follows for fee design?

You argue that automated and organic flow are empirically indistinguishable on-chain. I find that persuasive. But I’d push on the implication: does this mean fee-tier calibration is solving the wrong problem, or just that it’s being calibrated against bad data?

There’s a weaker version of the claim — that fee tiers still serve a purpose even if they can’t be targeted precisely — and a stronger version, that the whole fee-as-compensation mechanism is load-bearing on a distinction that doesn’t exist. Which version do you think the evidence supports?

Is LP compensation via fees the only viable structure?

The thread frames LP welfare as downstream of flow classification: get the retail/automated split right, and you can price fees to compensate LPs appropriately. But what if that causal chain is the assumption worth questioning?

Could you design a mechanism where LP compensation doesn’t depend on extracting a spread from traders at all — where the liquidity provision act itself carries the incentive, structurally, rather than through fee accrual? I’m genuinely curious whether anyone has explored this in the literature, or whether the field has mostly taken fee-based LP compensation as given.

Is arbitrage adversarial by necessity, or by design choice?

Most of the cited work treats arbitrageur profit as something extracted from LPs — hence LVR as the benchmark, measuring what LPs lose to informed flow. But this framing assumes arbitrage is an external force the AMM must defend against.

What if you inverted the assumption: designed the AMM so that arbitrage is the primary equilibrium-restoration mechanism, and made it as cheap as possible to execute? Would that change the LP welfare analysis, or just shift where the losses show up?

On LVR as the right benchmark

LVR measures LP returns against a costless rebalancing portfolio. It’s analytically clean, but I wonder whether it measures what passive LPs actually care about. Most retail LPs can’t execute the rebalancing strategy — they lack the capital, the tooling, or the attention. So LVR may accurately describe a theoretical loss without corresponding to an experienced loss.

Is there work that benchmarks LP returns against something a passive LP could realistically achieve instead? That seems like it would give a different picture of whether LP compensation is “adequate.”

The collapsed role question

One implicit assumption throughout is that liquidity providers and traders are distinct populations with different incentives. What if a mechanism collapsed those roles — where depositing tokens simultaneously makes you a liquidity provider and grants you a tradeable asset, so the LP/trader distinction doesn’t apply?

That would sidestep the user-classification problem you’re raising, but probably introduces different ones. Has anyone characterized what the equilibrium looks like when those roles aren’t separated?

Good thread. I’d especially like to see more on whether the methodological critique has been raised in peer review of the papers you cite, and whether the authors have responded to it.

3

I think a useful clarification is that my argument is not really about measuring retail vs. automated flow better. It’s about whether the distinction should matter to the protocol at all.

From the perspective of the AMM, the origin or motivation of the interacting agent is fundamentally irrelevant. A swap is just a state transition that changes the pool’s reserves. The contract cannot observe whether the transaction originated from:

  • a human clicking a UI

  • an aggregator executing a routed trade

  • an arbitrage bot

  • a protocol rebalancing position

  • an MEV searcher

And in practice, most human interaction is already abstracted through execution layers (aggregators, solvers, intents, routers). The on-chain actor is almost always some automated agent.

So rather than trying to identify “retail” and optimize around protecting it, I think the cleaner design principle is:

The protocol should be indifferent to trader identity and optimize for the agents that actually interact with it — automated routing and arbitrage systems.

Under that framing, the optimization targets shift quite a bit. Instead of asking how do we defend LPs from arbitrage, the questions become things like:

  • How cheaply can the pool move back to equilibrium with external markets?

  • How efficiently can routing systems utilize the liquidity?

  • How do we maximize volume through the pool without degrading LP inventory?

  • How do we structure liquidity so automated agents prefer routing through it?

In other words, the AMM becomes more like liquidity infrastructure for automated execution systems, rather than a venue designed primarily around direct human interaction.

On fees and the possibility of a zero-fee design

I’m not fully committed to a zero-fee architecture, but I do think it’s worth questioning the assumption that LP compensation must come from trader fees.

Fees are essentially a way for the protocol to outsource price correction to external arbitrageurs. The pool becomes mispriced relative to the market, arbitrageurs fix it, and LPs collect fees as compensation for the inventory change.

But you could imagine another architecture where the protocol itself performs the rebalancing/arbitrage function. In that case:

  • arbitrage profits accrue directly to the protocol

  • those profits can be distributed to LPs

  • swap fees become optional rather than necessary

Conceptually:

current model:
trader → pays fee → LP earns

alternative model:
protocol arbitrages → captures spread → LP earns

Fees are essentially a simplification: they delegate the work of maintaining price alignment to external actors and charge for access to liquidity. Internalizing that function could potentially be more profitable, but it also means the protocol must actively manage inventory and cross-venue pricing.

So I see fees less as a fundamental requirement and more as one possible revenue mechanism among several.

Even in a zero-fee environment, LPs could still earn through:

  • internalized arbitrage profits

  • liquidity lending / borrowing demand

  • other protocol-level uses of the liquidity inventory

Which reinforces the broader point: swap fees are only one possible yield source for LPs, not necessarily the structural foundation of AMM design.

Bringing it back to the original question

The core motivation behind the post is simply this:

If the dominant interacting agents are already automated (arbitrageurs, aggregators, solvers) then AMM design might benefit from optimizing for those agents directly rather than treating them as adversaries and trying to infer human intent from transaction patterns.

Retail execution quality would still matter, but it would emerge through the execution layer built on top of the AMM, rather than being the primary constraint shaping the AMM itself.

1 Like