1% paper/sections/literature.tex — UQO WP 27, "The Invisible Hand of Machines"2% Section: Related literature. Every citation key is in paper/references.bib and every claim is3% backed by the corresponding row of docs/literature_ledger.md.4\section{Related literature}5\label{sec:literature}67The paper sits at the intersection of six literatures: agent-based and heterogeneous-agent models of8financial markets; the microstructure of double auctions and the measurement of price discovery;9information aggregation with correlated signals, herding and feedback trading; liquidity spirals,10fire sales and the measurement of contagion; algorithmic trading, flash crashes, algorithmic11monoculture and algorithmic collusion; and, closest to our question, the rapidly growing work that12places large language model (LLM) agents in economic experiments and market simulations. We review13each in turn and close with a statement of what the present paper adds and where it overlaps with14existing work.1516\subsection{Agent-based and heterogeneous-agent models of financial markets}17\label{sec:literature:abm}1819Our baseline population of rule-based traders descends from agent-based computational finance, which20showed that interacting heterogeneous rules reproduce the fat tails, volatility clustering and bubbles21of real markets \citep{lebaron2006}. In the Santa Fe artificial stock market the speed of belief22updating governs whether prices track fundamentals or display excess volatility and technical-trading23regimes \citep{lebaron1999}; in adaptive belief systems, discrete-choice switching between predictors24produces increasingly complex price dynamics as the intensity of choice rises \citep{brock1997}, and25the share of trend followers determines whether the fundamental price is stable \citep{brock1998}.26Interacting noise traders who switch between optimistic, pessimistic and fundamentalist states generate27scaling laws and clustered volatility endogenously \citep{lux1999,lux2000}; herd behaviour among28speculators yields bubbles and crashes in a mean-field model of contagious optimism \citep{lux1995}; and29recruitment-type opinion dynamics generate persistent herding states without any informational reason30\citep{kirman1993}. These mechanisms matter for us because they describe what a \emph{classical}31population does on its own, before any LLM is introduced, and because trend following and herding are32precisely the channels through which a correlated LLM order flow can be amplified.3334Closer to our exchange design, \citet{chiarella2002} simulate a double auction populated by35fundamentalist, chartist and noise traders and reproduce order-book regularities, and36\citet{chiarella2009} show that heterogeneous rules in a limit-order book generate fat tails,37volatility clustering and order-flow autocorrelation, with the chartist share driving excess38volatility. Leverage in an agent-based model produces fat tails and clustered volatility through39synchronized margin-driven selling \citep{thurner2012}, which is the closest antecedent of our leverage40treatment. Multi-agent reinforcement-learning markets reproduce the same stylized facts41\citep{lussange2021} and are a precursor of LLM-populated markets. The surveys of \citet{hommes2006}42and \citet{hommes2021} summarize the laboratory and agent-based evidence that expectation heterogeneity43drives aggregate instability; we follow their lead in treating an LLM policy as one more expectation44type. \citet{farmer2009} argue that agent-based models are the appropriate tool for out-of-equilibrium45dynamics and for policy counterfactuals that cannot be solved analytically; \citet{axtell2025}46identify validation, calibration and scale as the open frontiers, which our matched-seed design and47power analysis address directly. For calibration and validation we use the moment set of structural48stochastic volatility models \citep{franke2012} and the catalogue of stylized facts: heavy tails49\citep{mandelbrot1963,cont2001}, conditional heteroskedasticity \citep{engle1982,bollerslev1986}, the50long memory of absolute returns \citep{ding1993}, and the long memory of order-flow signs coexisting51with nearly efficient prices \citep{lillo2004}.5253\subsection{Double auctions, zero intelligence and the measurement of price discovery}54\label{sec:literature:lob}5556The continuous double auction is the institution we study, and its experimental pedigree is long:57\citet{smith1962} showed that double auctions with induced values converge to competitive equilibrium58with few traders, and \citet{smith1988} showed that experimental asset markets with known dividend59streams nevertheless produce bubbles and crashes; our dividend-paying asset and bubble metrics follow60that design. The zero-intelligence tradition established that market structure, not trader61rationality, drives allocative efficiency \citep{gode1993}, a point reviewed by \citet{ladley2012}. A62zero-intelligence model of the continuous double auction predicts spread and price impact from63order-flow rates alone \citep{farmer2005}, and the statistical theory of \citet{smith2003} ties64spread, depth and price diffusion to the same parameters. These results fix the benchmark against65which any effect of ``intelligence'' must be measured: a change in the composition of a population can66move liquidity merely by changing order-flow rates, which is why we report realized order and volume67shares alongside the assigned treatment. On the empirical side we rely on the order-book facts of68\citet{bouchaud2002}, the Markovian book model of \citet{cont2010}, the linear relation between69order-flow imbalance and short-horizon price changes \citep{cont2014}, the survey of70\citet{gould2013}, and the treatment of impact and the slow digestion of order flow in71\citet{bouchaud2018}. Our event-driven exchange, with per-agent latency and a price--time matching72engine, follows the design of ABIDES \citep{byrd2020}, and our realism checklist is that of73\citet{vyetrenko2020}.7475The outcome variables are standard measures of price discovery and market quality. The price-impact76coefficient $\lambda=\Cov(v,Y)/\Var(Y)$ of \citet{kyle1985} and the adverse-selection spread of77\citet{glosten1985} tie liquidity to the information content of order flow. \citet{hasbrouck1991}78measures that content by the permanent price impact in a VAR of quote revisions and trades;79\citet{hasbrouck1993} decomposes the transaction price into an efficient random-walk component and a80pricing error whose standard deviation is a market-quality metric; \citet{hasbrouck1995} attributes81price discovery across venues through information shares, which we adapt to LLM versus rule-based82order flow. Variance ratios reject the random walk under autocorrelation \citep{lo1988}, the effective83spread can be inferred from serial covariance \citep{roll1984}, and illiquidity can be measured as84absolute return per unit of volume \citep{amihud2002}. \citet{bai2016} define informativeness by how85well prices forecast future fundamentals, and \citet{bond2012} argue that informativeness is the86welfare-relevant object because real decisions feed on prices. \citet{hendershott2014} separate87transitory price pressure from permanent information effects, which justifies our separation of the88pricing error from the efficient-price variance. Because the simulator observes the latent value, we89can compute these statistics against the truth rather than against a proxy, which is the main90advantage of a simulated economy for this question \citep{fama1970}.9192\subsection{Information aggregation with correlated signals, herding and feedback}93\label{sec:literature:info}9495The theoretical core of the paper is the aggregation of \emph{correlated} private signals.96\citet{hellwig1980} shows that competitive prices aggregate dispersed private signals imperfectly97because of noise supply, and \citet{grossman1980} that prices cannot be fully informationally98efficient when information is costly. With strategic informed traders, \citet{foster1996} show that99the correlation structure of signals, not only their precision, governs trading intensity and price100informativeness, and \citet{back2000} that more correlated information leads to faster revelation.101\citet{goldstein2015} show that information diversity across traders is complementary, so that a loss102of diversity reduces informativeness about some components of value; \citet{dugast2018} show that103abundant cheap data can lower informativeness by crowding out costly deep information, a close parallel104to shared-model crowding; and \citet{kondor2012} warn that public information can raise disagreement105through higher-order beliefs, so that shared inputs do not trivially homogenize behaviour.106\citet{pasquariello2007} ties information heterogeneity among strategic speculators to contagion and107shows that homogeneity reduces it. Our stylized model (Section~\ref{sec:theory}) borrows the108correlated-error structure of these papers and reinterprets the common component as the signature of a109shared foundation model or prompt family.110111Herding supplies the behavioural counterpart. Sequential decision-makers who follow their predecessors112destroy information aggregation \citep{banerjee1992}, informational cascades become more likely as113signals become more correlated \citep{bikhchandani1992}, reputational concerns make managers mimic one114another \citep{scharfstein1990}, and short-horizon speculators herd on the same information source115even when it is unrelated to fundamentals \citep{froot1992}. The buy-side imbalance statistic of116\citet{lakonishok1992} is the standard empirical measure of institutional herding, and we compute it117for LLM cohorts. Feedback trading closes the loop: the share of trend followers determines the118stability of the fundamental price in heterogeneous-belief models \citep{brock1998}, and119contagious optimism among speculators produces bubbles and crashes \citep{lux1995}. Our feedback120extension (Proposition~\ref{prop:feedback}) is a mechanical version of this amplification.121122\subsection{Liquidity spirals, fire sales, leverage and contagion measurement}123\label{sec:literature:spirals}124125The leverage treatment and the H3 and H6 hypotheses draw on the literature on funding constraints.126\citet{brunnermeier2009} show that market liquidity and funding liquidity reinforce each other through127margins, producing liquidity spirals; our margin-call mechanism follows this logic. Liquidation values128fall when potential buyers are themselves constrained \citep{shleifer1992}, constrained arbitrageurs can129amplify rather than correct mispricing \citep{shleifer1997}, forced sales push prices below130fundamentals and propagate losses \citep{shleifer2011}, and financially constrained arbitrageurs may131under-provide liquidity \citep{gromb2002}. Endogenous margins amplify booms and crashes in the leverage132cycle, with leverage caps as a policy lever \citep{geanakoplos2010}, and procyclical intermediary133leverage amplifies price moves \citep{adrian2010}. In the agent-based setting, \citet{thurner2012}134obtain fat tails and clustered volatility from synchronized margin-driven selling. Across assets,135losses of convergence traders propagate through reduced risk-bearing capacity \citep{kyle2001},136distressed selling induces endogenous correlation \citep{cont2013}, fire-sale spillovers scale with137common exposures and leverage \citep{greenwood2015}, cross-market rebalancing transmits shocks between138assets with unrelated fundamentals \citep{kodres2002}, and dense networks are robust to small shocks139but fragile to large ones \citep{acemoglu2015}. For measurement we use the spillover index of140\citet{diebold2009}, its generalized directional form \citep{diebold2012} and its network141interpretation \citep{diebold2014}, and we heed the warning of \citet{forbes2002} that correlation-based142contagion tests must adjust for heteroskedasticity.143144\subsection{Algorithmic trading, flash crashes, monoculture and algorithmic collusion}145\label{sec:literature:algo}146147The evidence on algorithmic and high-frequency trading is, on average, favourable to liquidity and148price discovery: algorithmic trading improves liquidity \citep{hendershott2011}, high-frequency traders149trade in the direction of permanent price changes and against transitory pricing errors150\citep{brogaard2014}, and algorithmic strategies in foreign exchange improve price efficiency even151though they are more correlated with each other than human trades \citep{chaboud2014}. That last152finding is the one we take most seriously: correlation among algorithms is a documented fact, and its153consequences depend on what the algorithms are correlated \emph{on}. High-frequency market makers154behave as new market makers with tight inventory control \citep{menkveld2013}, a feature we build into155the market-maker policy; \citet{menkveld2016} reviews the speed, adverse-selection and liquidity156consequences, and \citet{foucault2016} show how fast traders reacting to common news signals affect157informativeness and liquidity. The darker evidence concerns stress. During the Flash Crash,158high-frequency traders did not trigger the crash but their synchronized inventory management amplified159it \citep{kirilenko2017}; order-flow toxicity rose before the crash and liquidity provision withdrew160when flow became toxic \citep{easley2011}; the August 2007 quant unwind revealed that common strategies161produce simultaneous deleveraging \citep{khandani2011}; and ultrafast extreme events proliferate in a162machine ecology operating beyond human response time \citep{johnson2013}. \citet{budish2015} show that163continuous-time trading creates a latency arms race with correlated sniping and propose frequent batch164auctions, which motivates our latency-matching treatment and call-auction robustness arm.165166The algorithmic-monoculture literature gives the homogenization mechanism a name. \citet{kleinberg2021}167show that when many decision-makers rely on the same algorithm, outcomes can be worse even if that168algorithm is individually more accurate; \citet{bommasani2021} identify homogenization as a169structural risk of foundation models, and \citet{bommasani2022} document outcome homogenization170across deployers of shared models. On the output side, generative AI raises individual creativity but171reduces collective diversity \citep{doshi2024}, standard LLM outputs show diminished diversity of172thought relative to human samples \citep{park2024brm}, and LLM assistance homogenizes human ideation173\citep{anderson2024}; conversely, ensembles of twelve LLMs match human crowd forecasting accuracy174\citep{schoenegger2024}, which illustrates the diversity benefit our H4 tests in a market. The175algorithmic-collusion literature shows that learning algorithms can coordinate without being told to:176Q-learning pricing algorithms learn tacit collusion \citep{calvano2020,klein2021}, algorithmic pricing177raised margins in German retail gasoline \citep{assad2024}, asymmetric adoption softens competition178\citep{brown2023}, LLM pricing agents reach supracompetitive prices with a sensitivity to prompt179wording \citep{fish2024}, and LLM sellers in continuous double auctions collude more when they can180communicate \citep{agrawal2025}. In securities markets, \citet{colliard2026} show that Q-learning market181makers learn adverse-selection pricing but quote markups that shrink with the number of algorithms,182and \citet{dou2025} show that reinforcement-learning informed speculators in a Kyle-type market sustain183supracompetitive profits through price-trigger strategies or homogenized learning biases, lowering184price informativeness and liquidity. \citet{dou2025} is the nearest theoretical neighbour of our185mechanism; the difference is that their channel is strategic, whereas ours is a shared-error structure186in signals, and that they have no language models.187188The policy literature has already drawn the systemic conclusion. \citet{gensler2020} argue that deep189learning increases uniformity and interconnectedness because models built on the same data generate190correlated predictions; \citet{danielsson2022} show that AI in risk management can create endogenous191systemic risk through common models and procyclical responses; the \citet{fsb2024} warns that common AI192models and data sources could raise correlations in trading, lending and pricing;193\citet{danielsson2025} model AI as amplifying existing vulnerabilities and making crises faster;194\citet{aldasoro2025} list monoculture and correlation among the financial-stability concerns of an195intelligent financial system; and \citet{meng2026} propose a unified model in which the AI adoption196share and signal correlation raise systemic coupling convexly, with 13F evidence of tail-loss197amplification. None of these provides a controlled experiment; the present paper is an attempt to198supply one.199200\subsection{LLM agents in economics and markets}201\label{sec:literature:llm}202203\paragraph{LLMs as simulated subjects.}204\citet{horton2023} proposed using LLMs as simulated economic agents to replicate classic behavioural205experiments at low cost; \citet{aher2023} showed that LLMs can simulate multiple humans and replicate206human-subject studies with fidelity that varies by model and task; and \citet{argyle2023} showed that207conditioning on demographic backstories produces ``silicon samples'' matching sub-population response208distributions. GPT choices in budgetary tasks score higher on rationality than humans and are largely209stable across demographic prompts \citep{chen2023pnas}; chatbot choices in behavioural games fall within210the human distribution and are more cooperative than the median human \citep{mei2024}; in repeated211games LLMs cooperate well but coordinate poorly \citep{akata2025}; and expectations elicited from an212LLM reading newspaper articles track professional forecasters and are extrapolative213\citep{bybee2023}. \citet{manning2024} automate hypothesis generation and testing with LLM subjects, and214\citet{immorlica2024} propose a framework in which generative agents are economic agents with new215primitives such as prompts and context. Generative agents with memory, reflection and planning produce216believable emergent social behaviour \citep{park2023}, social-network simulations reproduce217information propagation \citep{gao2023s3}, and EconAgent produces realistic macroeconomic dynamics218\citep{li2024econagent}. \citet{anthis2025} conclude that LLM social simulations are a promising219method provided that validity concerns---homogeneity, sycophancy, drift---are addressed; homogeneity is220the concern this paper turns into a treatment.221222\paragraph{Single-agent trading and backtests.}223ChatGPT-derived sentiment on headlines predicts next-day returns \citep{lopezlira2023chatgpt}.224FinMem \citep{yu2023finmem}, TradingGPT \citep{li2023tradinggpt} and FinAgent \citep{zhang2024finagent}225evaluate one LLM agent, or a small team with layered memory and distinct characters, by backtesting on226historical prices. In these designs the market does not react to the agent, so nothing can be learned227about equilibrium prices, and \citet{lopezlira2025memorization} show that such backtests are228contaminated by memorization of the historical period. Both points motivate a synthetic economy with229an endogenous price and no history to memorize.230231\paragraph{LLM-populated market simulators.}232A second generation places many LLM agents on a matching engine. ASFM couples LLM traders to a real233order-matching system and finds reactions to rate and inflation shocks consistent with economic priors234\citep{gao2024asfm}; StockAgent adds bulletin-board communication and reports that GPT- and235Gemini-driven agents differ systematically in trading behaviour \citep{zhang2024stockagent};236TwinMarket scales to a social-media-connected population, reproduces bubbles and crashes, and reports237that agents of one model family are more homogeneous \citep{yang2025twinmarket}; FCLAgent combines238LLM-decided direction with rule-based sizing and reproduces path-dependent patterns absent from239conventional agent-based models \citep{hashimoto2025}; StockSim provides an open order-level simulator240with latency, slippage and price--time matching \citep{papadakis2025}; and \citet{li2026switch} find241that LLM agents' trading-style switching in a year-long simulation is only partially consistent with242behavioural-finance theory. \citet{lopezlira2025trade} is the closest in spirit: LLM agents in a243persistent order-book simulation adhere to assigned strategies, generate price discovery and bubbles,244and are analysed with partial-dependence-style perturbations. These papers establish that LLM markets245\emph{can} produce the patterns we study. They are descriptive in the sense that the population246composition is fixed, the information set of LLM agents is not held equal to that of the agents they247replace, and the fundamental value, the efficient price and the transaction price are not separated.248249\paragraph{Experimental-finance replications.}250A third strand replicates laboratory markets with LLM subjects. In Smith--Suchanek--Williams-type251markets, LLM agents price near fundamentals and form only muted bubbles, with lower variance in their252strategy text than humans \citep{henning2025}. In a simulated call auction, LLM agents exhibit253disposition effects and extrapolative beliefs, and prompt interventions causally change bubble size254\citep{ouyang2026}. Across fifteen models, \citet{saxena2026} find behaviour ranging from fundamental255pricing to human-like bubbles, with mixed populations unstable and advanced models adapting to exploit256weaker ones. Replications of \citet{smith1962} find slower or absent convergence with model-family257differences \citep{struski2026}, and the trading institution alone moves the efficiency of identical258LLM agents from 56\,\% to 89\,\% \citep{chupilkin2026}. These studies answer the question of259behavioural fidelity---whether LLMs behave like humans---and their answers conflict. Our results do not260require fidelity: they require only that errors be correlated within a model.261262\paragraph{Correlated behaviour from shared inputs.}263The single paper that measures correlated trading from shared inputs is \citet{song2026}: overlap in264retrieved documents among retrieval-augmented trading agents raises mean pairwise action correlation265from 0.04 to 0.41 and increases position volatility without raising profits. That design is a backtest266without market feedback, so it cannot speak to informativeness or liquidity. \citet{cho2025} document267peer conformity in LLM multi-agent systems, with the confidence gap and format of peer information268controlling its strength. In LLM re-creations of herding experiments, \citet{hansen2025} find that AI269agents rely more on private signals than humans but can be prompted to herd when maximizing profit.270Finally, LLM service behaviour drifts over months \citep{chen2024hdsr}, which is why we cache every271response and replay from a frozen decision library.272273\subsection{What this paper adds}274\label{sec:literature:adds}275276We should be candid about overlap. The hypothesis that shared models correlate trading is not ours: it277is stated by \citet{gensler2020}, the \citet{fsb2024} and \citet{danielsson2025}, modelled as278homogenized learning biases by \citet{dou2025}, and measured in a backtest by \citet{song2026}.279Model-family homogeneity inside LLM markets has been observed by \citet{yang2025twinmarket},280\citet{zhang2024stockagent} and \citet{saxena2026}. LLM-populated order-book simulators exist and are281open source \citep{lopezlira2025trade,papadakis2025,gao2024asfm}. Our simulator is therefore not the282contribution; the design is. First, we treat the LLM penetration share as a randomized treatment283assigned to trader slots by an independent seed stream, so that matched pairs across the share share284every other stream---fundamentals, news, endowments, latencies, policy assignment, activation,285tie-breaks---and differences in outcomes are causal effects on the same market history286(Section~\ref{sec:identification}). Second, we separate three prices: the latent value known to the287simulator only, the public value implied by announced dividends and news, and the transaction price, so288that an information-efficiency loss can be distinguished from transitory pressure289\citep{hasbrouck1993,hendershott2014}. Third, we derive the shared-error mechanism in a Kyle--Hellwig290setting with a model-specific common error and an effective number of models, and then test it by291intervention---prompt diversity at equal share, a two-model mix, inherited information sets, memory,292communication cohorts---so that the effective number of models is manipulated rather than observed.293Fourth, we report the realized dose (the share of orders and volume actually originated by LLM agents)294alongside the assigned share, and we run fairness designs with equal decision opportunities, equal295dollar budgets and matched latency, because naive comparisons confound ``being an LLM'' with being slow296and infrequent. Fifth, the leverage interaction and the safeguard counterfactuals connect the LLM297literature to \citet{brunnermeier2009}, \citet{thurner2012} and \citet{kirilenko2017}, with thresholds,298estimators and seed counts fixed before the first confirmatory call.299