Competitive Market Behavior of LLMs

2609.02580 ↗ · 2026-09-02 · cs.MA, cs.AI, econ.GN

Pawel Struski, Jakub Swistak, Inez Okulska, Przemyslaw Biecek

The finding

Headline claim

Markets populated by LLM agents fail to converge to competitive equilibrium in Smith-style double auctions, with worse price dispersion than human subjects, though one smaller model shows partial convergence.

The test

Replication of Smith (1962) continuous double auction experiments with 11 buyers and 11 sellers (equilibrium price $2.00, quantity 6), replacing humans with three LLM populations (GPT Large, GPT Small, Gemini Large), 5 trading rounds of up to 300 iterations, 10 runs per condition, compared against the original human benchmark on price dispersion (alpha), trade volume, and allocative efficiency.

Effect size & scope

No condition fully converged in any round; best condition (GPT Small) reached alpha of 11.7 in round 5 versus the human benchmark of 3.5, roughly 3x human dispersion, with mean price $0.13 above equilibrium and efficiency peaking at 0.91 versus full efficiency; Gemini Large and GPT Large oscillated with no downward trend, with efficiency as low as 0.36 (GPT Large round 4) and trades as low as 2.2 versus equilibrium 6.

Limitations

The paper concedes convergence is partial rather than absent, differs from prior negative findings (Jia and Yuan 2024) partly due to implementation choices, and only one of three models reproduced the human pattern. Unstated: tiny scale (one market design, 5 rounds, 10 seeds, 3 models, temperature 1), prompt sensitivity (agents given profit-maximization but no trade-off guidance), single unspecified good with stylized valuations, no analysis of whether more rounds or different prompts close the gap, and results section truncated so CoT analysis claims are not verifiable from this text.

Counterpoint

A skeptic would note that 10 runs of one highly stylized market with default temperature and arbitrary prompt framing is far too thin to generalize about LLMs and market institutions; the finding that the smallest model converges most could reflect prompt artifacts or truncation/verbosity limits rather than capability; and the CoT lexical analysis risks over-reading token-level signals. The result may say more about prompt engineering than about markets.

Economy implication

Little or nothing yet for the economics of AI at deployment scale. The relevant signal is directional: if LLM agents deployed as autonomous economic actors (pricing, trading, negotiation) do not reliably reach the outcomes market institutions assume, that is a small early data point for market-design and regulatory thinking, but this lab setting cannot yet support any claim about real trading or pricing systems.

Who this matters for

Technical Capability: This measures a concrete agentic failure mode, sequential strategic decision-making in an institutionally constrained environment, that standard benchmarks do not test, and shows bigger models can do worse.

Regulatory & Legal: It complicates the emerging collusion-in-pricing-agents debate, since supracompetitive or non-convergent behavior by LLM traders may reflect general incompetence rather than collusive intent.

Market & Valuation: If financial exchanges ever run at speeds where LLM agents trade autonomously, these results suggest the allocative efficiency that justifies double-auction market design cannot be assumed to survive that substitution.

Abstract

Large language models (LLMs) are increasingly deployed as economic agents, yet there is little evidence whether LLM agents are suited for participating in market mechanisms designed for humans, and whether these mechanisms deliver desired outcomes when faced with LLM agents. We address this question by replicating seminal economic experiments, replacing human subjects with LLM agents. We place agents in a double auction environment, which is a widely-used market mechanism. We check whether such a market is able to deliver an efficient allocation of resources, thereby testing a novel dimension of alignment of LLM agents -- their compatibility with a fundamental market mechanism. We find that markets populated by LLM agents exhibit slower or no convergence towards market equilibrium, thus providing less efficient allocations than markets populated by humans. We then analyze agents' individual trading decisions and find substantial heterogeneity both across model families and market roles. We also run a lexical analysis of Chain-of-Thought (CoT) traces generated by the agents. We find that the decision to execute a trade rather than continue incrementally adjusting prices is associated with a shift from strategic considerations toward urgency. We publicly release our testing framework, which can be used for future evaluations.