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 draws on six literatures: agent-based models of financial markets; double-auction8microstructure and the measurement of price discovery; information aggregation with correlated9signals, herding and feedback; liquidity spirals, fire sales and contagion; algorithmic trading,10monoculture and algorithmic collusion; and, closest to our question, LLM agents in economic experiments11and market simulations. We close with what this paper adds and where it overlaps with existing work.1213\subsection{Agent-based and heterogeneous-agent models}14\label{sec:literature:abm}1516Our rule-based population descends from agent-based computational finance, in which interacting17heterogeneous rules reproduce fat tails, volatility clustering and bubbles \citep{lebaron2006}. In the18Santa Fe artificial stock market the speed of belief updating governs whether prices track fundamentals19or display excess volatility and technical-trading regimes \citep{lebaron1999}; in adaptive belief20systems, switching between predictors produces increasingly complex dynamics as the intensity of choice21rises \citep{brock1997}, and the share of trend followers determines whether the fundamental price is22stable \citep{brock1998}. Noise traders switching between optimistic, pessimistic and fundamentalist23states generate scaling laws and clustered volatility endogenously \citep{lux1999,lux2000}, contagious24optimism yields bubbles and crashes \citep{lux1995}, and recruitment dynamics generate persistent25herding without any informational reason \citep{kirman1993}. These are the channels through which a26correlated LLM order flow can be amplified by a classical population. Closer to our exchange,27\citet{chiarella2002} reproduce order-book regularities with fundamentalist, chartist and noise traders28in a double auction, and \citet{chiarella2009} show that heterogeneous rules in a limit-order book29generate fat tails, volatility clustering and order-flow autocorrelation, with the chartist share30driving excess volatility. Leverage in an agent-based model produces fat tails and clustered volatility31through synchronized margin-driven selling \citep{thurner2012}, the closest antecedent of our leverage32treatment, and multi-agent reinforcement-learning markets reproduce the same facts33\citep{lussange2021}. Following the surveys of \citet{hommes2006} and \citet{hommes2021}, we treat an34LLM policy as one more expectation type; following \citet{farmer2009} and \citet{axtell2025}, we use35the agent-based method for counterfactuals that cannot be solved analytically and address its open36frontiers---validation, calibration, scale---with matched seeds and a power analysis. Validation37targets are the moment set of \citet{franke2012} and the stylized facts of heavy tails38\citep{mandelbrot1963,cont2001}, conditional heteroskedasticity \citep{engle1982,bollerslev1986} and the39long memory of absolute returns \citep{ding1993}.4041\subsection{Double auctions, zero intelligence and price-discovery measurement}42\label{sec:literature:lob}4344\citet{smith1962} showed that double auctions with induced values converge to competitive equilibrium45with few traders, and \citet{smith1988} that experimental asset markets with known dividend streams46nevertheless produce bubbles and crashes; our dividend-paying asset and bubble metrics follow that47design. Zero-intelligence traders constrained by budgets reach near-full allocative efficiency, so48market structure rather than rationality drives efficiency \citep{gode1993,ladley2012}, and a49zero-intelligence model of the double auction predicts spread and price impact from order-flow rates50alone \citep{farmer2005,smith2003}. This fixes the benchmark against which any effect of51``intelligence'' must be read: changing the composition of a population can move liquidity merely by52changing order-flow rates, which is why we report realized order and volume shares alongside the53assigned treatment. The book is parametrized from the empirical regularities in \citet{bouchaud2002}, \citet{cont2010},54\citet{cont2014}, \citet{gould2013} and \citet{bouchaud2018}; the event-driven exchange with per-agent55latency follows ABIDES \citep{byrd2020} and the realism checklist is that of \citet{vyetrenko2020}.5657The outcomes are standard measures of price discovery. The price-impact coefficient58$\lambda=\Cov(v,Y)/\Var(Y)$ of \citet{kyle1985} and the adverse-selection spread of \citet{glosten1985}59tie liquidity to the information in order flow; \citet{hasbrouck1991} measures that information by the60permanent price impact in a VAR of quotes and trades, \citet{hasbrouck1993} decomposes the transaction61price into an efficient random-walk component and a pricing error, and \citet{hasbrouck1995}62attributes price discovery through information shares, which we adapt to LLM versus rule-based flow.63Variance ratios test the random walk \citep{lo1988}, the effective spread follows from serial64covariance \citep{roll1984}, and illiquidity is absolute return per unit of volume \citep{amihud2002}.65\citet{bai2016} define informativeness by how well prices forecast future fundamentals,66\citet{bond2012} argue that it is the welfare-relevant object because real decisions feed on prices,67and \citet{hendershott2014} separate transitory price pressure from permanent information effects,68which justifies separating the pricing error from the efficient-price variance. Because the simulator69observes the latent value, these statistics are computed against the truth rather than a proxy, which70is how we operationalize the efficiency notions of \citet{fama1970}.7172\subsection{Correlated signals, herding and feedback}73\label{sec:literature:info}7475The theoretical core of the paper is the aggregation of \emph{correlated} private signals. Competitive76prices aggregate dispersed signals imperfectly because of noise supply \citep{hellwig1980} and cannot77be fully efficient when information is costly \citep{grossman1980}. With strategic informed traders,78the correlation structure of signals, not only their precision, governs trading intensity and79informativeness \citep{foster1996}, and more correlated information leads to faster revelation80\citep{back2000}. Information diversity across traders is complementary, so a loss of diversity81reduces informativeness about some components of value \citep{goldstein2015}; abundant cheap data can82lower informativeness by crowding out costly deep information \citep{dugast2018}, a close parallel to83shared-model crowding; public information can raise disagreement through higher-order beliefs, so84shared inputs do not trivially homogenize behaviour \citep{kondor2012}; and information heterogeneity85among speculators generates contagion while homogeneity reduces it \citep{pasquariello2007}. Our86stylized model (Section~\ref{sec:theory}) borrows this correlated-error structure and reinterprets the87common component as the signature of a shared foundation model. Herding supplies the behavioural88counterpart: sequential decision-makers who follow predecessors destroy information aggregation89\citep{banerjee1992}, cascades become more likely as signals become more correlated90\citep{bikhchandani1992}, reputational concerns make managers mimic one another91\citep{scharfstein1990}, and short-horizon speculators herd on the same information source even when it92is unrelated to fundamentals \citep{froot1992}; the buy-side imbalance statistic of93\citet{lakonishok1992} is the standard measure and we compute it for LLM cohorts. Feedback trading94closes the loop, since the share of trend followers governs the stability of the fundamental price95\citep{brock1998}; Proposition~\ref{prop:feedback} is a mechanical version of this amplification.9697\subsection{Liquidity spirals, fire sales and contagion}98\label{sec:literature:spirals}99100Hypotheses H3 and H6 draw on the literature on funding constraints. Market liquidity and funding101liquidity reinforce each other through margins, producing liquidity spirals \citep{brunnermeier2009};102our margin-call mechanism follows this logic. Liquidation values fall when potential buyers are103themselves constrained \citep{shleifer1992}, constrained arbitrageurs can amplify rather than correct104mispricing \citep{shleifer1997} and may under-provide liquidity \citep{gromb2002}, forced sales push105prices below fundamentals and propagate losses \citep{shleifer2011}, endogenous margins amplify the106leverage cycle with leverage caps as a policy lever \citep{geanakoplos2010}, and procyclical107intermediary leverage amplifies price moves \citep{adrian2010}. Across assets, losses propagate through reduced risk-bearing capacity \citep{kyle2001}, distressed108selling induces endogenous correlation \citep{cont2013}, fire-sale spillovers scale with common109exposures and leverage \citep{greenwood2015}, rebalancing transmits shocks between unrelated assets110\citep{kodres2002}, and dense networks are fragile to large shocks \citep{acemoglu2015}. For measurement we use the spillover index of \citet{diebold2009}, its111generalized directional form \citep{diebold2012} and network interpretation \citep{diebold2014}, and we112heed the warning of \citet{forbes2002} that correlation-based contagion tests must adjust for113heteroskedasticity.114115\subsection{Algorithmic trading, monoculture and algorithmic collusion}116\label{sec:literature:algo}117118The average evidence on algorithmic trading is favourable: it improves liquidity119\citep{hendershott2011}, high-frequency traders trade with permanent price changes and against120transitory errors \citep{brogaard2014}, and algorithmic strategies in foreign exchange improve price121efficiency even though they are more correlated with each other than human trades122\citep{chaboud2014}---the finding we take most seriously, since correlation among algorithms is a123documented fact whose consequences depend on what they are correlated \emph{on}. High-frequency124market makers keep tight inventory control \citep{menkveld2013}, a feature of our market-maker policy;125\citet{menkveld2016} reviews the speed and adverse-selection consequences and \citet{foucault2016} show126how fast traders reacting to common news affect informativeness and liquidity. Under stress the picture darkens: during the Flash Crash, high-frequency traders did not127trigger the crash but their synchronized inventory management amplified it \citep{kirilenko2017};128order-flow toxicity rose and liquidity withdrew when flow became toxic \citep{easley2011}; the August1292007 quant unwind showed common strategies deleveraging simultaneously \citep{khandani2011}; and130ultrafast extreme events proliferate in a machine ecology beyond human response time131\citep{johnson2013}. \citet{budish2015} show that continuous trading creates a latency arms race with132correlated sniping and propose frequent batch auctions, which motivates our matched-latency treatment133and call-auction robustness arm.134135Algorithmic monoculture names the homogenization mechanism: when many decision-makers rely on the same136algorithm, outcomes can be worse even if it is individually more accurate \citep{kleinberg2021};137foundation models carry a homogenization risk because downstream systems inherit the same failure modes138\citep{bommasani2021}, and shared models mis-serve the same individuals across deployers139\citep{bommasani2022}. Generative AI raises individual creativity but reduces collective diversity140\citep{doshi2024}, standard LLM outputs show diminished diversity of thought \citep{park2024brm}, and141LLM assistance homogenizes human ideation \citep{anderson2024}; conversely, ensembles of twelve LLMs142match human crowd forecasting accuracy \citep{schoenegger2024}, the diversity benefit that H4 tests in a143market. Learning algorithms also coordinate without being told to: Q-learning pricing algorithms learn144tacit collusion \citep{calvano2020,klein2021}, algorithmic pricing raised retail gasoline margins145\citep{assad2024}, asymmetric adoption softens competition \citep{brown2023}, LLM pricing agents reach146supracompetitive prices with a sensitivity to prompt wording \citep{fish2024}, and LLM sellers in double147auctions collude more when they can communicate \citep{agrawal2025}. In securities markets, Q-learning148market makers quote markups that shrink with the number of algorithms \citep{colliard2026}, and149reinforcement-learning speculators in a Kyle-type market sustain supracompetitive profits through150price-trigger strategies or homogenized learning biases, lowering informativeness and liquidity151\citep{dou2025}. \citet{dou2025} is the nearest theoretical neighbour of our mechanism: their channel is152strategic, ours is a shared-error structure in signals, and they have no language models. The policy153literature has drawn the systemic conclusion: deep learning increases uniformity because models built154on the same data generate correlated predictions \citep{gensler2020}; common AI models and data sources155could raise correlations in trading, lending and pricing \citep{fsb2024}; AI amplifies existing156vulnerabilities and makes crises faster \citep{danielsson2025}; monoculture and correlation are among157the stability concerns of an intelligent financial system \citep{aldasoro2025}; and the AI adoption158share and signal correlation raise systemic coupling convexly, with 13F evidence of tail-loss159amplification \citep{meng2026}. None of these provides a controlled experiment.160161\subsection{LLM agents in economics and markets}162\label{sec:literature:llm}163164\paragraph{Simulated subjects.}165\citet{horton2023} proposed LLMs as simulated economic agents for replicating behavioural experiments166at low cost; \citet{aher2023} showed that they can simulate multiple humans with fidelity varying by167model and task, and \citet{argyle2023} that demographic backstories yield silicon samples matching168sub-population responses. GPT choices score higher on rationality than humans and are stable across169demographic prompts \citep{chen2023pnas}, chatbot choices in behavioural games fall within the human170distribution \citep{mei2024}, LLMs cooperate well but coordinate poorly in repeated games171\citep{akata2025}, and LLM-elicited expectations are extrapolative \citep{bybee2023}. Generative agents172with memory and reflection produce believable emergent behaviour \citep{park2023,gao2023s3}, EconAgent173produces realistic macroeconomic dynamics \citep{li2024econagent}, and \citet{immorlica2024} treat174generative agents as economic agents with prompts as primitives \citep[see also][]{manning2024}.175\citet{anthis2025} conclude that LLM social simulations are promising provided that homogeneity,176sycophancy and drift are addressed; homogeneity is the concern this paper turns into a treatment.177178\paragraph{Single-agent trading and backtests.}179ChatGPT-derived headline sentiment predicts next-day returns \citep{lopezlira2023chatgpt}. FinMem180\citep{yu2023finmem}, TradingGPT \citep{li2023tradinggpt} and FinAgent \citep{zhang2024finagent}181evaluate one agent, or a small team with layered memory, by backtesting on historical prices. The182market does not react to the agent, so nothing is learned about equilibrium prices, and such backtests183are contaminated by memorization of the historical period \citep{lopezlira2025memorization}---two184reasons for a synthetic economy with an endogenous price and no history to memorize.185186\paragraph{LLM-populated market simulators.}187A second generation places many LLM agents on a matching engine. ASFM finds reactions to rate and188inflation shocks consistent with economic priors \citep{gao2024asfm}; StockAgent adds bulletin-board189communication and reports that GPT- and Gemini-driven agents differ systematically190\citep{zhang2024stockagent}; TwinMarket scales to a social-media-connected population, reproduces191bubbles and crashes, and reports that agents of one model family are more homogeneous192\citep{yang2025twinmarket}; FCLAgent reproduces path-dependent patterns absent from conventional193agent-based models \citep{hashimoto2025}; StockSim is an open order-level simulator with latency and194price--time matching \citep{papadakis2025}; and trading-style switching in a year-long simulation is only195partially consistent with behavioural-finance theory \citep{li2026switch}. \citet{lopezlira2025trade} is196closest in spirit: LLM agents in a persistent order book adhere to assigned strategies, generate price197discovery and bubbles, and are analysed with partial-dependence-style perturbations. These papers show198that LLM markets \emph{can} produce the patterns we study; they hold the population composition fixed,199do not equalize the information set of LLM agents with that of the agents they replace, and do not200separate the fundamental value, the efficient price and the transaction price.201202\paragraph{Experimental-finance replications.}203In Smith--Suchanek--Williams-type markets, LLM agents price near fundamentals and form only muted204bubbles, with less variable strategy text than humans \citep{henning2025}; in a call auction they205exhibit disposition effects and extrapolative beliefs, and prompt interventions causally change bubble206size \citep{ouyang2026}; across fifteen models, behaviour ranges from fundamental pricing to human-like207bubbles, mixed populations are unstable, and advanced models adapt to exploit weaker ones208\citep{saxena2026}; replications of \citet{smith1962} converge more slowly or not at all, with209model-family differences \citep{struski2026}; and the institution alone moves the efficiency of210identical agents from 56\,\% to 89\,\% \citep{chupilkin2026}. These studies ask whether LLMs behave211like humans and their answers conflict. Our results do not require fidelity, only that errors be212correlated within a model.213214\paragraph{Correlated behaviour from shared inputs.}215The one paper that measures correlated trading from shared inputs is \citet{song2026}: overlap in216retrieved documents among retrieval-augmented trading agents raises mean pairwise action correlation217from 0.04 to 0.41 and increases position volatility without raising profits. That is a backtest without218market feedback, so it cannot speak to informativeness or liquidity. \citet{cho2025} document peer219conformity in LLM multi-agent systems, and \citet{hansen2025} find in re-created herding experiments220that AI agents rely more on private signals than humans but can be prompted to herd when maximizing221profit. LLM service behaviour drifts over months \citep{chen2024hdsr}, which is why every response is222cached and replayed from a frozen library.223224\subsection{What this paper adds}225\label{sec:literature:adds}226227We should be candid about overlap. The hypothesis that shared models correlate trading is not ours: it228is stated by \citet{gensler2020}, the \citet{fsb2024} and \citet{danielsson2025}, modelled as229homogenized learning biases by \citet{dou2025}, and measured in a backtest by \citet{song2026}.230Model-family homogeneity inside LLM markets has been observed by \citet{yang2025twinmarket},231\citet{zhang2024stockagent} and \citet{saxena2026}, and open LLM order-book simulators exist232\citep{lopezlira2025trade,papadakis2025,gao2024asfm}. Our simulator is not the contribution; the design233is. First, the LLM penetration share is a randomized treatment assigned to trader slots by an234independent seed stream, so that matched pairs share every other stream---fundamentals, news,235endowments, latencies, policy assignment, activation, tie-breaks---and paired differences are causal236effects on the same market history (Section~\ref{sec:identification}). Second, we separate three237prices---the latent value known to the simulator only, the public value implied by announced dividends238and news, and the transaction price---so that an efficiency loss can be distinguished from transitory239pressure \citep{hasbrouck1993,hendershott2014}. Third, we derive the shared-error mechanism in a240Kyle--Hellwig setting with a model-specific common error and an effective number of models, then test241it by intervention---prompt diversity at equal share, a two-model mix, inherited information sets,242memory, communication cohorts---so that the effective number of models is manipulated rather than243observed. Fourth, we report the realized dose, the share of orders and volume actually originated by244LLM agents, alongside the assigned share, and run fairness designs with equal decision opportunities,245equal dollar budgets and matched latency, because naive comparisons confound being an LLM with being246slow and infrequent. Fifth, the leverage interaction and safeguard counterfactuals connect the LLM247literature to \citet{brunnermeier2009}, \citet{thurner2012} and \citet{kirilenko2017}, with thresholds,248estimators and seed counts fixed before the first confirmatory call.249