Competitive Market Behavior of LLMs
Pawel Struski, Jakub Swistak, Inez Okulska, Przemyslaw Biecek
cs.MA · cs.AI · econ.GN
Quick verdict
Claims LLM traders don't converge to double-auction equilibrium as humans do; solid benchmarked replication but only 3 models, 10 runs, no significance tests.
Summary
The paper replicates Vernon Smith's 1962 double-auction experiments, replacing all 22 human traders with LLM agents (GPT Large, GPT Small, Gemini Large) in a continuous double auction with a persistent order book. Each of 10 independent runs per model comprises 5 rounds; 11 buyers and 11 sellers hold reservation prices from $0.75 to $3.25, giving a competitive equilibrium of $2.00 and 6 trades. Against Smith's human benchmark, no LLM market fully converges: price dispersion (α) stays above the human value in every round, and only GPT Small reproduces the human pattern of steadily declining dispersion (27.6 to 11.7), while Gemini Large and GPT Large oscillate. GPT Large's aggressive rent-seeking ($0.01 improvements, reluctance to cross spreads) suppresses trade volume to 2.2–3.8 per round and efficiency to 0.36–0.67. The authors document role asymmetries (GPT Small buyers make larger bid improvements and enjoy higher fill rates), opening-offer anchoring, and a lexical analysis of Gemini's reasoning traces linking spread-crossing to urgency vocabulary. They argue these results caution against reading supracompetitive LLM prices as collusion and against uncritically using LLMs as human stand-ins in ABMs. Main limitations: only three models from two providers, 10 runs per condition, means reported without dispersion or significance tests, a possibly binding 300-iteration round cap, and reasoning analysis limited to one model.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Markets populated by the three tested LLM populations never fully converge to the competitive equilibrium ($2.00, q=6), with price dispersion α above the Smith (1962) human benchmark in every trading round.
- GPT Small is the only condition reproducing the human pattern of steadily declining dispersion (27.6→11.7) and is the most efficient market despite being the least capable model.
- GPT Large's reluctance to cross spreads (persistent $0.01 improvements) suppresses trade volume to 2.2–3.8 per round and allocative efficiency to 0.36–0.67.
- In Gemini Large's reasoning traces, the decision to cross the spread coincides with a lexical shift from strategy/optimization terms toward urgency/execution terms.
- Because non-colluding LLM traders do not reliably converge to competitive equilibrium, supracompetitive prices in LLM markets cannot be attributed cleanly to collusion.
Strengths
Canonical design with sharp dual benchmarks
Smith (1962) symmetric baseline: 11 buyers/11 sellers, reservation prices $0.75–$3.25, equilibrium exactly ($2.00, q=6), so allocative efficiency and dispersion α are precisely measurable against both theory and decades of human-subject data (§3.1, Table 1).
Materially better mechanism than prior replication
A persistent order book admitting only spread-improving quotes with immediate execution on crossing (§3.2) fixes the passive price-posting design of Jia and Yuan (2024), and 10 seeds per condition address LLM stochasticity where prior work ran single runs.
Honest epistemic hedging
Actionable cross-domain implications
Full artifact release
Weaknesses
No uncertainty quantification at market level
Narrow external validity vs. generalizing title
GPT Large results confounded with the T=300 cap
Post-hoc, unvalidated lexical categories
Benchmark commensurability assumed, not shown
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether LLM agents substituted for human traders in Vernon Smith's (1962) continuous double auction reproduce the famously robust human result of convergence to competitive equilibrium. Three claims carry it: (1) no tested LLM population fully converges — price dispersion α stays above the human benchmark in every round (Table 1); (2) behavior is heterogeneous across model families and buyer/seller roles (§4.2, Figures 2, 3, 5); (3) spread-crossing decisions in Gemini's reasoning traces shift lexically from strategy/optimization terms to urgency/execution terms (§5, Figure 4). A derived claim (§6) is methodological: non-convergence by non-colluding LLMs weakens collusion inferences in the Fish et al. / Agrawal et al. line. The market-institution alignment framing is apt and novel, but with three models from two providers, the title's generalization to 'LLMs' outruns the evidence.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main finding of 'Competitive Market Behavior of LLMs'?
When LLM agents replace the human traders of Vernon Smith's classic 1962 double-auction experiments, the markets fail to fully converge to the competitive equilibrium: price dispersion stays above the human benchmark in every round, and only the smallest tested model (GPT Small) shows the human-like pattern of steadily declining dispersion.
Which LLMs were tested, and how many simulation runs were used?
Three models: gpt-5.4-2026-03-05 (GPT Large), gpt-5.4-mini-2026-03-17 (GPT Small), and gemini-3.1-pro-preview (Gemini Large). Each condition consists of 10 independent simulated markets of 5 trading rounds each, with 11 buyers and 11 sellers per market.
Why does this paper matter for algorithmic-collusion research?
Inferring collusion from above-equilibrium prices presumes that honest, non-colluding agents would converge to the competitive equilibrium. The paper shows LLM traders do not reliably converge even without collusion, so supracompetitive prices alone are weak evidence of collusion in LLM markets.
Is the code and data from the paper available?
Yes. The simulation framework, full prompt templates, and experimental data are publicly released at github.com/jswistak/competitive-market-simulation, so Table 1's numbers can be re-derived from the released runs and new models can be evaluated with the same harness.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The typical-experiment walkthrough that §5 says is 'in the supplementary material' is not part of the provided text; all main tables/figures (1–6) and appendices A–D are present.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.