At a Glance
LLM-driven traders do not reliably recreate human market behavior in a basic double-sided auction: prices stay more scattered and markets leave gains on the table, and the smallest model tested was the closest to human-like performance.
ON THIS PAGE
What They Found
LLM agents trading in a classic double auction failed to fully converge to the competitive price and quantity that human traders typically reach. Price variation around the theoretical equilibrium stayed consistently higher than the human benchmark across rounds, and total realized surplus never reached full efficiency. A smaller model produced the most human-like pattern, while larger models often stalled with incremental undercutting rather than closing trades. human benchmark
By the Numbers
1Human price dispersion fell from 11.8% in round 1 to 3.5% in round 5, while all LLM conditions kept dispersion well above the human benchmark in every round.
2Best-performing model peaked at 0.91 allocative efficiency (91% of maximum possible surplus); no model achieved full efficiency (1.00).
3Competitive equilibrium target was price $2.00 and quantity 6 trades per round; every condition fell short of 6 trades in nearly every round (only one exception: the small model in round 5).
What This Means
Engineers building marketplaces or autonomous trading agents should validate agent behavior against human and theoretical benchmarks before deployment, because off-the-shelf large models may not produce efficient market outcomes. Platform and policy designers assessing market fairness, collusion risk, or agent governance should treat supracompetitive prices and incremental undercutting cautiously—these patterns can arise even without collusion. market fairness
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1: Market design. Reservation prices induce symmetric step-function demand and supply schedules. The competitive equilibrium ( p ⋆ , q ⋆ ) = ( $ 2.00 , 6 ) (p^{\star},q^{\star})=(\$2.00,\,6) lies at their intersection.
Fig 2: Figure 2: Empirical Cumulative Distribution Function (ECDF) of order improvement size for GPT Small agents across simulations (pooled). Small order improvements dominate among sellers, while buyers tend to place more large improvements.
Fig 3: Figure 3: The first attempted rent (profit) for agents whose side is first to place an initial order and second (individual observations). Left panel: GPT Large. In all but one case, the 2nd moving agent immediately falls to his reservation price. Right panel: GPT Small. The pattern is partly reversed.
Fig 4: Figure 4: Lexical shift in the internal thinking between move types, based on terms distinctively associated with urgency, execution, optimization, or strategy.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
The study covers three model populations from two providers and was not designed to isolate the effect of model size, so the surprising better performance of the smallest model is suggestive but not conclusive. Only one model exposed internal reasoning traces, so the analysis of decision language (urgency versus margin protection) is illustrative, not general proof. Experiments had limited rounds and a single symmetric market setup; different horizons, information, or market rules could change outcomes. internal reasoning traces
Methodology & More
Used a classic laboratory market setup where 11 buyers and 11 sellers each have private reservation prices spread from $0.75 to $3.25, creating symmetric supply and demand that intersect at a competitive price of $2.00 and an equilibrium quantity of 6 trades per round. Simulated continuous double auctions with three LLM populations across 10 simulations per condition and measured trade count, transaction prices, price dispersion relative to $2.00, and allocative efficiency (realized surplus relative to maximum). Human benchmark data from the original Smith experiments served as the reference for convergence. Found that none of the LLM conditions consistently reached the competitive equilibrium: price dispersion stayed higher than in human experiments, trade volumes usually fell short of 6, and allocative efficiency never hit 100% (best model reached 91%). Behavioral analysis of one model that exposed internal reasoning showed two phases: cautious incremental improvements focused on protecting profit margins, and occasional larger jumps tied to urgency and execution to close trades. Practical implications include a warning against using LLMs as plug-and-play human proxies in agent-based market models and caution when interpreting high prices or tiny repeated undercuts as evidence of collusion—those patterns can reflect model-specific dynamics rather than coordinated wrongdoing. The authors release a simulation framework to test more models, market designs, and longer horizons to clarify whether incomplete convergence reflects a fundamental limitation or a slower path to equilibrium. simulation framework longer horizons
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Authors are affiliated with recognized research institutions/universities (NASK, Warsaw U. of Technology, Adam Mickiewicz), though citation/h-index signals are limited. Venue is arXiv, so solid but not top-tier by provided signals.