betaFeedback

WorksComputer Science

The Market Is the Reward: Design Markets, Not Just Traders for Human-AI Collective Intelligence

Abstract

Learning agents are becoming market participants, yet machine learning still treats the market as a fixed environment, although its payoff rule is the reward agents learn from. This position paper argues that financial AI should build HyperMarkets for human-AI collective intelligence, and not only the agents that trade in them: markets whose liquidity provision, information production, outcome determination, and trading rule are performed or adapted by learning agents and designed jointly toward explicit market-level objectives. In a pre-registered simulation with Claude agents trading for principals, joint design raises a market-level objective by 79% over separately tuned functions, through a trading rule that helps only when slow traders are present, and how evidence is paid for decides whether a stronger agent earns more for its principal. We state when fixed rules or non-market aggregation should be preferred and what findings would refute the case.

Introduction

Learning agents are becoming market participants. Agents built on foundation models read filings, use tools, and hold positions ([72, 74, 67, 73]), and reinforcement learning now fine-tunes a language model directly on trading rewards ([71]). In forecasting, large language models can extract return-relevant information from financial news and predict subsequent stock returns ([41]), a retrieval-augmented language-model system approaches the accuracy of aggregated competitive human forecasters ([27]), and in a three-month tournament an ensemble of 12 language models was statistically indistinguishable from a crowd of 925 human forecasters ([57]). Agents already trade on people’s behalf, and what they achieve depends on the market and on the agent (Figure 2a,Figure 2b; [61, 33]). Machine learning has approached these agents as it approaches most agents: train or prompt a policy, place it in an environment taken as given, and report its score ([39, 8]). Financial benchmarks have grown from static tasks ([68, 69, 48]) to agentic workflows ((Li et al., 2024; Cao et al., 2026; Peng et al.,

Figure 2. Agents already trade for people, the market decides what they achieve, and complementarity is unproven. (a) In Project Swap, agents traded books for 201 people under fixed rules; 85% of the shortfall came from representing preferences ([33]). (b) In Project Deal, the same bike sold for $38 by Haiku and $65 by Opus, while participants rated the fairness of the two models’ deals 4.06 and 4.05 on a 1–7 scale ([61]). (c) Across 106 experiments, human–AI combinations performed worse on average than the better of humans or AI alone ([62]).

  1. and to live trading across markets ([53]), yet each takes the market as the fixed backdrop against which an agent is scored. In a market, this setup leaves out most of what determines the score.

A market is an unusual environment for a learner, because its dynamics are produced by other learners and by a rule. Q-learning pricing algorithms in a repeated oligopoly consistently learn supracompetitive prices without communicating ([10]), learning algorithms with bounded rationality can sustain any feasible and individually rational payoff ([12]), and reasoning language-model agents tacitly collude in pricing games even when instructed otherwise ([54]). When many decision-makers converge on one algorithm, overall decision quality can fall even if that algorithm is more accurate for each of them alone ([36]), and models that share components can homogenize outcomes across deployments ([6]). A market’s payoff rule is also the reward from which agents trained in it learn: in continuous limit order books, price correlations break down at millisecond horizons, and the resulting mechanical arbitrage is something competition only makes faster, so what the market pays for is speed ([7]). Holding the market fixed therefore fixes, without discussion, both the environment and the reward for a growing class of agents. A return obtained in such a market belongs to the agent and the market together, yet it is credited to the agent.

What has changed is that the rest of the market can now be learned as well. A market runs on a few functions: liquidity provision, the production of information that prices respond to, the determination of how claims resolve, and the rule that matches orders. Market design has engineered these functions around human participants [56]: people whose attention is scarce, whose strategizing is costly, and whose signals are trusted because they are costly to fake. Each function is now a learning problem. Liquidity provision is a sequential decision problem addressed with online learning and reinforcement learning [1, 60]. Information production is forecasting and generation [27], and in an online labor market generative AI has already weakened the signaling value of written applications [23]. Outcome determination is evidence assessment; it is still done by people and remains open [22]. The rule itself can be learned, as in learned auctions and economic policies trained against adaptive agents [18, 75]. For the first time, every function of a market is something machine learning can build, and all of them act on one shared environment.

We therefore argue that financial AI should build HyperMarkets for human–AI collective intelligence, and not only the agents that trade in them. A HyperMarket is a market in which one or more core functions are performed or adapted by learning agents, designed jointly with its human and AI participants toward explicit market-level objectives (Figure 1). Complementarity between people and agents is likewise something to demonstrate (Figure 2c).

Agent-centred research and HyperMarket research

Figure 1. What is held fixed, and what is designed. Left: a common agent-development setup, in which one trained agent (ringed square) acts in a market whose functions, population, and objective are given, and its return is credited to it alone. Right: HyperMarket research treats the market’s functions as designed or learned (blue), serving people (orange circles) and agents (green squares) toward a declared market-level objective. Grey marks what is held fixed.

The shift changes the research problem in three ways. The object becomes a system of interacting learners and a learnable rule: a multi-agent, bilevel problem rather than a single policy. The objective moves to the level of the market, such as how informative its prices are and whether the people trading in it are served rather than exploited; a misspecified market-level objective can be gamed as surely as any reward [59]. Evaluation changes with them: a result should report the market in which it was obtained, including the mechanism, the population, and shared base models. Prediction markets are our starting case because their prices are public and their claims resolve, which makes collective intelligence measurable against the world; each function also has a counterpart in securities markets.

This paper makes four contributions. It reframes the market, for a learning agent, as both environment and reward, so that market design becomes part of the learning problem. It gives a function-by-function account of what learning changes in liquidity provision, information production, outcome determination, and the trading rule, each tied to existing evidence. It states falsifiable criteria, grounded in a formal statement of the design problem as a bilevel program, naming the comparisons that would support the position and the findings that would weaken it, and runs one such comparison: a pre-registered simulation with Claude agents, in which joint design plays through an interaction we did not predict while the channel we did predict fails. And it proposes concrete actions for researchers, benchmark builders, venues, and operators.

Sections 2–6 develop these points in order. We advocate a research direction; the simulation illustrates the proposed evaluation and does not test any deployed market.

The Market Belongs in the Learning Problem

Environment, reward, and credit

For an agent that learns in a market, three things that are usually separate coincide. The environment is non-stationary because other participants learn too. The reward is the market’s payoff rule. The data are prices that other agents’ actions produced. Multi-agent learning treats co-learners as part of the environment, and two-level formulations learn a policy for the environment’s designer alongside the agents’ policies [75]. Environments that respond to deployed models are studied as performative prediction [50] and strategic classification [30]; a market whose participants learn is such an environment, with the added feature that its rule is also a reward. Financial-agent research rarely adopts this view: the market is inherited and held fixed, and the score credits the agent with an outcome it shares with its market, much as people explain behavior by the person rather than the situation [55].

A market also does more than combine answers that already exist. It creates incentives to acquire information, determines when participants can act on it, allocates exposure, and specifies how claims become payments. An ensemble receives a collection of forecasts; a market can influence which forecasts are worth producing. Information acquisition must be rewarded, which is why prices can reveal only part of the available information [26], and strategic aggregation depends on the structure of the market [45].

The functional perspective on financial systems separates what must be done from the institution that happens to do it [43, 42]. We organize the learning problem around liquidity provision, information production, outcome determination, and the trading rule. Price discovery results from these functions and participation, rather than being a fifth component to optimize separately. The same four functions exist in securities markets under other names. Dealers and designated market makers provide liquidity; analysts, rating agencies, and data vendors produce information; clearing houses settle trades, and for credit derivatives industry committees decide whether a credit event occurred; exchanges set matching rules, which can be redesigned, as frequent batch auctions show [7]. Prediction markets make the argument easiest to state because their claims resolve cleanly, but the questions carry over: which of these functions learning agents now perform, what they are rewarded for, and how the people on the other side are protected. Securities markets, with their regulation and systemic risk, are the next setting rather than the first.

An agent’s objective and a market’s objective are distinct. A dealer could earn more by widening spreads while making information costly to express; a forecaster could be accurate but redundant with every other forecaster; an automated resolver could be fast while excluding evidence it cannot process. Choosing the rule that governs everyone’s orders lies outside a trader’s action space; the proposed shift adds such choices to what financial AI builds.

The design problem. Let θ\theta collect the designer’s choices (quoting policy, evidence rewards, settlement procedure, clearing rule), restricted to a set ΘC\Theta_{\mathrm{C}} that encodes fixed commitments such as contract terms, exposure limits, update schedules, and dispute rights. Given θ\theta, people h∈Hh \in H and agents a∈Aa \in A choose or learn policies whose joint response π∗(θ)\pi^{*}(\theta) is an equilibrium or the limit of the learning dynamics. The research problem is

max⁡θ∈ΘCJ(θ,π∗(θ))s.t.Uh(θ,π∗(θ))≥Uh∀h∈H,(1)\max_{\theta\in\Theta_C} J(\theta,\pi^*(\theta)) \quad\text{s.t.} \quad U_h(\theta,\pi^*(\theta)) \ge U_h \quad\forall h \in H, \tag*{(1)}

where JJ is the declared market-level objective, for example the proper-score accuracy of prices net of subsidy and review costs, and each person’s outcome must not fall below a reference level UhU_h, such as their outcome in the existing market. Agent-centered research fixes θ\theta and improves one agent’s policy. Treating θ\theta as a variable is what makes the problem bilevel, and the inner response π∗(θ)\pi^*(\theta) is what makes it a learning problem rather than a static design.

What is new, and what is not

Market design already treats institutions as objects to be engineered [56]. Automated market makers provide rule-based liquidity [29, 3], convex market making connects that problem to online learning [1], and learned auctions and two-level policy learning [18, 75] build on automated mechanism design [16], so learning a mechanism is an established idea. What is new is an agent that strategizes as fully as a rule rewards and that can perform market functions rather than only bid. HyperMarket research extends and connects these lines of work for such agents. Financial AI evaluation has followed the same path from models to agents (Figure 4): it began with static tasks, moved to tool-using workflows and simulated firms, and now runs agents in live markets. Each step widened what the agent does while holding the market fixed. The next step is to make the market’s functions the object of design. A recent central-bank review calls the design of rules for markets populated by AI agents an important and largely unstudied question [13]; auction format and platform rules already change whether algorithms collude [5, 35].

Figure 4. What financial AI evaluation varies. Benchmarks by what they score, dated by first public version: static tasks ([68, 69, 38, 48]), agent workflows ([37, 28, 11, 49]), simulated markets ([8, 39]), and live markets ([53]). In each, the market is held fixed.

Joint design means that each function is designed and evaluated with the responses of the others and of the participants in view, as the inner response π∗(θ)\pi^*(\theta) expresses, rather than that one model controls every function or that human preferences become training parameters. Learning is also compatible with stable rules, which ΘC\Theta_C encodes: a policy can be trained offline, tested, and held fixed for a contract’s lifetime, with adaptation confined to declared parameters and update intervals. An ordinary exchange with an LLM trader falls outside our definition, since only the trader learns.

What a HyperMarket should be better at

A HyperMarket might expand the range of claims it can sustain under a stated budget, the quality of information its prices convey, and the conditions under which people can contribute and benefit. A claim of progress must identify the capability gained, the costs paid, and whose outcomes changed.

Building the Functions of a HyperMarket

Each function in Figure 3 was designed around a limit of human participants that learning agents relax. Each subsection gives the evidence on that limit, the learning problem the function becomes, and the design commitment it requires, and ends with a result from one pre-registered simulation (Appendix A). In it, a sponsor with a fixed budget funds monitoring, liquidity, and settlement review for up to 60 dataset-release contracts, traded for 12 days by principals represented by strong or weak agents and by slow traders. We report 40 held-out worlds, and six in which Claude Sonnet and Haiku 4.5 read the news.

Figure 3. Each market function, before and after learning agents. Each column reads top to bottom: the human limit the function was designed around, the learning problem it becomes, the design commitment § 3 argues for, and what our simulation found. The functions act on one shared environment: evidence informs quotes, quotes shape orders, and resolved outcomes feed evaluation.

Liquidity provision: make information actionable

Prediction markets remain thin on all but prominent questions, because human liquidity follows human attention ([65, 4]), and information moves a price only if someone can trade on it at acceptable cost. The logarithmic market scoring rule offers bounded-loss automated quoting ([29]), and learned dealers treat quoting as sequential decision-making under inventory risk ([60]), and language models can now serve as the policies of hierarchical trading learners ([24]). The HyperMarket question is how to direct that learning toward informative prices at a declared subsidy, rather than toward the dealer’s profit. A release-date market could spread a fixed support budget over a contract’s life according to evidence arrival, with a fixed-liquidity market maker as comparator. Quotes should adapt to declared public signals, not to inferred human vulnerability, and the learned action set must preserve the scoring rule’s loss bound.

In simulation, liquidity is the function to fix first: making it ideal adds 0.37 to the market objective JJ when done first and nothing when done last, while better information adds 0.09 first and 0.33 last (Figure 5). A component evaluated alone is misvalued.

Bar chart showing gains in the market objective when liquidity, information, trading rule, or settlement is improved first or last

Figure 5. Liquidity first. Gain in the market objective JJ when one function of the fixed design is made ideal, improved first (dark) or last (light); 40 held-out worlds.

Information production: pay for evidence, not text

Language-model forecasters already rival human crowds, and averaging human and machine forecasts was more accurate still ([57]); retrieval improves their reading of financial time series ([66]). Evidence from equity markets suggests that AI adoption can change the informational environment in which trading agents operate: Lopez-Lira and Tang [41] show that GPT-4 predicts returns from firm-specific news, but this predictability declines as LLM adoption expands. As AI reduces information-processing frictions, the value of processing certain types of information may decline, changing incentives for information acquisition and trading. This motivates treating market mechanisms and information-production incentives as endogenous to the capabilities and adoption of AI agents. Yet markets have trusted signals that are costly to fake, which generative text weakens ([23]). Trading early on information that will be disclosed anyway earns private profit with little social value, unlike discovering what would otherwise stay hidden ([32]), and no current market mechanism aligns the two ([13]). The binding constraint can also be what agents know about their principals: when agents traded books for 201 people, participants ended at 0.55 on their own rankings against an attainable 0.89, and 85% of the shortfall came from imprecise representation of preferences ([33]); in financial advice, what investors chose and what serves their stated risk preferences can diverge ([64]). The learning problem is to reward the incremental contribution of evidence, accounting for cost, duplication, and strategic withholding, rather than the volume of plausible text; peer prediction ([44]) and the Bayesian truth serum ([51]) reward reports without ground truth. Agreement among copies of one agent counts as one source, and shared foundation models make such dependence a first-order concern ([6]; [36]).

In simulation, paying the monitor per report floods the record with re-reports and raises mis-settled contracts from 0.80 to 2.22 per world, in 34 of 40 worlds (Figure 6a). With Claude agents, the same choice decides who gains from a better agent. When the monitor is paid for new evidence, principals represented by Sonnet earn 12.5 ± 2.6 more per episode than those represented by Haiku in all six worlds, a gap of 0.67 standard deviations of individual profit that few would notice, as Project Deal found; when it is paid per report, the gap is 2.9 ± 3.9 (Figure 6b).

Figure 6. What the monitor is paid for. (a) Mis-settled contracts per world, 40 stylized worlds; red lines rise. (b) Profit of principals represented by Sonnet minus Haiku in six worlds with Claude agents reading the news, with the gap in standard deviations of individual profit.

Outcome determination: connect claims to the world

Settlement has been performed by people, and crowdsourced and decentralized oracle designs study how to organize it ([22]; [9]). Agents can gather and organize evidence while people adjudicate contested interpretations, but checking claims against structured filings and verifying each reasoning step remain open problems for current models ([63]; [70]). The learning problem is evidence assessment with abstention under adversarial inputs: interested participants may publish misleading evidence or place instructions in documents, so resolvers must be tested against fabricated and conflicting sources [40, 34], with trading and adjudication separated. The dataset contract should specify before trading what counts as a release and which sources are admissible, and a challenge window determines when a proposed outcome becomes final. Some contracts create incentives to cause the event itself [14] and should stay unlisted.

In simulation, we expected evidence to substitute for review, so that tuning monitoring and review together would pay more as evidence mattered more for settlement. It did not: that pair beats separate tuning only when evidence is irrelevant to settlement (Figure 7), because a reviewer tuned for the worst case still benefits from evidence when it arrives.

Figure 7. The coupling we predicted. JJ as evidence becomes more useful for settlement; each design re-tuned at each γ\gamma; bands are 95% intervals.

Trading rules: rich enough for agents, safe for people

Trading rules have been kept simple largely because strategizing is costly for people: eBay moved from auctions to posted prices [19, 21], and school districts adopted strategy-proof assignment partly to protect sincere families from sophisticated ones [46]. Learning agents make richer mechanisms usable, and in a mixed market they are the sophisticated players. Items sold by a stronger model fetched $3.64 more than when sold by a weaker one, yet participants rated the fairness of the two models’ deals almost identically [61]. Time is part of the problem: continuous trading rewards speed [7], and on decentralized exchanges ordering rules let bots extract value from other users [17]. Investors already vote with their orders: on Euronext Amsterdam, the share of non-proprietary volume in large-cap stocks traded on the continuous order book fell from 65% to under 45% between 2020 and 2025 as trading moved to the closing auction [13]. Periodic batch clearing is one response; Feeding America’s daily sealed-bid rounds let food banks without staff to watch auctions compete [52]. Participants must be able to anticipate and understand the rule, so updates belong between contracts, on a public schedule.

In simulation, uniform-price batch clearing brings prices closer to the attainable posterior and turns slow traders’ average result from −3.4-3.4 to +9.2+9.2, but only when slow traders are present; with agents alone the rules coincide (Figure 8). With Claude agents, whose beliefs are more moderate, batch clearing still raises JJ but no longer protects slow traders (−0.83±0.63-0.83 \pm0.63).

Figure 8. Batch clearing in a mixed market. (a) Distance of price from the posterior attainable from all evidence, with (solid) and without (dashed) slow traders. (b) Profit per episode, continuous (grey) and batch (colour); bars span the 10th–90th percentiles, circles are means.

Joint design pays through interactions between functions

Tuning all four functions jointly for the market objective, rather than each for its own, raises JJ from 2.62 to 4.69 under the same budget (paired difference 2.07 ±\pm 0.22; 37 of 40 worlds; Figure 9a), and batch clearing raises JJ with Claude agents as well (+1.09 ±\pm 0.27, six of six worlds; Figure 9b). The gain does not come from the channel we predicted (§ 3.3). It comes from a rule whose best setting depends on who is in the market, among functions whose values depend on one another. These results show that such interactions exist and can be measured, not their size in any real market.

Figure 9. Joint design under one budget. (a) JJ per world, functions tuned separately or jointly; red lines are worlds where joint design loses; circles are means with 95% intervals. (b) JJ per world with Claude agents under continuous (grey) and batch (blue) trading.

Joint design can also create dangerous dependencies: if one operator profits from trading and controls the outcome label, optimizing its total reward can undermine truthful settlement. Independent adjudication and separated access remain constraints even when a unified model would be cheaper; the objective is coordination among accountable functions rather than consolidation of authority in one learner.

Participation is not yet complementarity

People can contribute local knowledge, different access to information, and judgments about what makes a claim meaningful. Agents can contribute continuous monitoring and cheap replication of some analyses. Since human–AI combinations have on average underperformed the better of their parts (Figure 2c), a demonstration of complementarity must compare the combination with strong alternatives ([62, 31]). Assistance also differs from market-level complementarity: a tool that improves each person’s forecast ([58]) may leave a population’s information unchanged if everyone receives the same advice.

For prediction tasks, a first operational target is the quality of market probabilities measured against independently resolved outcomes with a declared proper scoring rule ([25]). Fairness, accessibility, and benefit to participants require separate measures: language-model credit scoring can help many applicants while biasing outcomes for some ([20]), and financial language models lag in Spanish and Greek relative to English ([74, 47]), which bears on who can take part in a market run through them.

What would support or weaken the position

The research should separate improvements in agents from improvements in institutions: hold agent capability and evidence access comparable while varying the mechanism, vary the population across human-only, AI-only, and mixed participation, and include a competitive non-market aggregator. All comparisons should report information, compute, subsidy, and human-time budgets, since a market that buys more evidence has found more evidence rather than aggregated it better. Figure 10 lists six tests and what our simulation settled. The position is weakened wherever gains vanish under matched budgets, the best design does not depend on the population, an ensemble matches the market at lower cost, or mixed participation worsens outcomes for people. Benefits to people must be measured on people, and a single success would establish possibility rather than broad applicability.

Figure 10. What would support or weaken the position. Each row is a test; each column a population. Our simulation settles some tests for stylized and Claude agents; none has been run with real people.

Alternative Views

Alternative Views

Improve and certify agents; leave institutions stable. Better forecasting, reliable tool use, and certification can address many failures without changing market infrastructure, and evidence of tacit collusion argues for certifying agents before they make market decisions [54]. The agendas are complementary: the first live agent marketplaces pair tests of agents with market rules for registration, rate limits, and failed deals [33]. But an agent’s behavior depends on the environment and population in which it acts, so better agents make better markets only under suitable rules. Adverse selection unravels a market precisely when buyers reason correctly about it [2, 21], so fully rational agents could thin some markets rather than improve them; in our simulation, the same agents gained or lost depending on the rule (§ 3.5). Certification also leaves open design questions such as how to reward information production or support an illiquid claim. Where certified agents under existing rules suffice for an objective, the existing rules should stay.

This is existing mechanism design with a new name. Learned auctions, adaptive market making, the AI Economist, and artificial prediction markets already cover parts of the proposal [18, 60, 75, 15]. The position concerns what financial AI should build and connect: agent-operated market functions studied together with the mixed populations they serve. Its value depends on finding interactions that isolated component studies miss; if the functions can be designed independently without consequential loss, the case for an integrated agenda is weaker.

Learning makes mechanisms less trustworthy. Fixed rules can be understood, audited, and anticipated, while a learned policy may be opaque, unstable, or exploitable through its updates. The proposal is compatible with offline learning behind fixed action constraints, human dispute procedures, and rules simple enough to follow; where a benefit fails to survive these restrictions, the fixed alternative should win.

Use ensembles or deliberation instead of markets. Forecasting ensembles already rival human crowds without budgets, strategy, or transaction costs. The case for a market is strongest when information is costly, distributed, and strategically supplied, and weakest when judgments can be collected cheaply from cooperative contributors.

Sufficiently capable agents make human participation unnecessary. People remain affected parties with purposes, rights, and information they may choose to withhold. Deciding whose interests count requires more than greater capability.

Call to Action

Researchers: make the mechanism a variable. A practical first project combines one learned function with fixed contracts and a strong baseline. State the intended market benefit before choosing the agent’s reward, and publish cases where fixed rules or non-market methods remain preferable. Later projects should study interactions between functions.

Benchmark and simulator builders: expose the market. Existing financial suites and leaderboards [69, 38] and agent simulators with budgets and delayed consequences [28, 53] are natural hosts for a market track. Provide separable interfaces for information acquisition, quoting, clearing, and settlement, including costs and disputes. Treat the population as a reported variable: results should state the mechanism, population, and shared base models, and be repeated across populations.

Venues and reviewers: recognize institutional progress. Invite results that improve a market even when its traders are unchanged, and ask of trading-agent results in which market they were obtained.

Operators: provide bounded, accountable test settings. Early studies should use simulation and consenting research participants before any claim about live deployment. Pilots should specify contracts, capital limits, and appeal procedures, and reports should separate better forecasting from transfers between participant groups.

Conclusion

Machine learning has made the market’s traders into learners. It can now make the rest of the market into learners as well: the functions that provide liquidity, produce information, settle claims, and set the rules. Holding those functions fixed while improving traders decides, without discussion, the environment and the reward for a growing class of agents. HyperMarkets make the alternative concrete: market functions performed or adapted by learning agents, designed with human and AI participants toward objectives stated at the level of the market. The opportunity is to sustain claims that human attention never reached and to use information markets have failed to elicit. The challenge is to show that the resulting coordination is worth its costs and serves the people in it. Human–AI collective intelligence is the objective to be established, not a benefit that follows from more capable agents.

Impact Statement

The proposed research concerns incentives and institutions that could affect people’s resources, information, and opportunities. Improved forecast accuracy does not justify extracting value from less informed participants, collecting unnecessary personal data, or creating harmful incentives around an event. Studies involving people require appropriate ethics review, informed consent, declared exposure limits, and mechanisms for disputes and withdrawal. Contracts whose incentives encourage harm should not be listed. This paper reports no new human-subject study or market deployment. Its proposed safeguards are design commitments, not evidence that a future implementation is safe.

AI Use Statement

OpenAI Codex and Anthropic Claude assisted with reviewing related literature, revising the manuscript, preparing figures and document-processing scripts, and implementing the simulation. In the simulation (Appendix A), Claude Sonnet and Haiku 4.5 act as experimental participants that read generated news and report probabilities; these outputs are simulation data, not evidence about people or live markets.

A Pre-Registered Simulation of Joint Design

The simulation illustrates the evaluation proposed in Section 4: the same budget, the same evidence, and the same participants, with only the design of the market’s functions varied. The hypotheses, metrics, and comparisons were written down before any run (docs/sim_prereg_zh.md); all are reported, including those that failed. Code, beliefs, and figures are in the supplementary material (sim/).

Environment

Claims and evidence. Each world has 60 candidate contracts of the form “will dataset XX be released by day 12?”. Each has a prior drawn from Beta(2, 2), a true outcome drawn from it, and an ambiguity level that makes settlement harder. Independent sources arrive at random days (Poisson, mean 3 per contract) with signed strength; each found source may be re-reported by aggregators (Poisson, mean 2), which adds no information.

Functions and design variables. Information: a monitor spends effort mm per contract and finds each source with probability 1−e−m1-e^{-m}; it is paid either per posted report, which makes posting re-reports profitable, or only for new independent evidence. Liquidity: an LMSR market maker with parameter bb and worst-case subsidy bln⁡2b\ln2. Settlement: review effort rr per contract; the chance of mis-settlement is aiexp⁡(−(reff/ρ+γEi))a_i\exp(-(r_{\mathrm{eff}}/\rho+\gamma E_i)), where aia_i is ambiguity, EiE_i the independent evidence found, and reffr_{\mathrm{eff}} the effort left after processing re-reports. γ\gamma is the evidence–settlement coupling we expected to drive the gains of joint design. Trading rule: continuous trading, in which agents act before slow traders within each step, or uniform-price batch clearing once per step. Every listed contract costs m+bln⁡2+rm+b\ln2+r from a shared budget of 400, and a contract is listed only if its expected mis-settlement rate is at most 5%.

Participants. Ten principals delegate to strong agents, which discount re-reports and combine public evidence with the principal’s private signal by Bayes’ rule; ten delegate to weak agents, which count re-reports as independent and read evidence with noise; ten slow traders hold only private signals and trade in 30% of steps. Each participant holds a private signal on a contract with probability 0.3.

Designs and objective. The fixed design uses hand-set defaults. In the separate design each function optimizes its own objective: the monitor its reports (a learned monitoring policy chooses m=3m=3 and posts re-reports), the market maker its loss (smallest bb), the reviewer the worst case without evidence, and the exchange its volume (continuous trading). The joint design maximizes the market objective JJ, the sum over listed contracts of the Brier improvement of the final price over the prior, minus 0.5 per mis-settled contract, by grid search subject to the slow traders’ average profit being at least −2-2. All tuning uses worlds 0–9; all reported numbers use worlds 100–139, paired across designs. Sensitivity runs vary re-report volume (0 or 4) and budget (250 or 600).

Pre-registered hypotheses and outcomes

Language-model agents

Evidence as news. For six held-out worlds (180 contracts), each found source was written as a dated news item (an official changelog, a maintainer’s statement) whose wording reflects its strength. Each re-report restates the same fact in different words, and half name an outlet; none repeats the identifier in the original item, so a careful reader can recognize but not trivially match it. This gives 481 independent items and 992 re-reports. The clean condition shows independent items only, corresponding to paying for new evidence; the echo condition adds the re-reports, corresponding to paying per report.

Elicitation and replay. Claude Haiku 4.5 (weak reader) and Claude Sonnet (strong reader) each read every contract under both conditions, 30 contracts per session, and reported a probability for each of the 12 days.

HypothesisOutcome (40 held-out worlds)
H1 Joint design wins, more so as γ\gamma growsPartly. +2.07 ± 0.22, 37 of 40 worlds; the gain of jointly tuning monitoring and review falls from +3.0 to −0.3 as γ\gamma rises from 0 to 1.
H2 Per-report pay induces re-reportsYes, largely by construction. Paying for new evidence cuts mis-settlement from 2.23 to 0.80; JJ moves from 2.62 to 2.71.
H3 Batch clearing protects slow tradersYes. Slow traders’ profit −3.4 → +9.2, JJ 2.53 → 4.69; no effect without slow traders.
H4 Decompose the shortfallOrder-dependent: information +0.09 first, +0.33 last; liquidity +0.37 first, −0.03 last; 52% of the gap remains with participants.
H5 Weak-agent principals lose, unnoticedDirection yes; gap/s.d. 0.85 (fixed), 1.39 (joint).

Table 1. Pre-registered hypotheses and what the stylized simulation found.

Instructions were neutral and did not mention re-reports. This produced 24 sessions and 720 belief paths. As in Project Deal and Project Swap ([61, 33]), beliefs were elicited once and replayed across all market designs, so differences between designs cannot come from sampling the models again. Principals’ private signals and slow traders are as in the stylized model; only reading the news is done by the models.

Results. Both models are roughly calibrated (Figure 11a). Re-reports change day-12 Brier scores only slightly (Sonnet 0.139 to 0.145, Haiku 0.151 to 0.150; Figure 11b), so the stylized assumption that weak agents double count does not hold for these models; why the gap shrinks under per-report pay (re-report timing, trading on belief paths) is not isolated here. Batch clearing raises JJ in both conditions (+1.03 ± 0.38 per report, five of six worlds; +1.09 ± 0.27 for new evidence, six of six; Figure 11c) but slightly lowers slow traders’ profit (−0.72 ± 0.42 and −0.83 ± 0.63): with the models’ more moderate beliefs, agents no longer trade far ahead of slow traders. The gap between principals represented by Sonnet and by Haiku is 12.5 ± 2.6 (continuous) and 11.5 ± 3.5 (batch) when the monitor is paid for new evidence, positive in all six worlds, and 2.9 ± 3.9 and 1.4 ± 4.6 when it is paid per report.

Figure 11. Claude agents, six held-out worlds, 180 contracts. (a) Share of contracts released by day-12 forecast bin. (b) Day-12 forecast with and without re-reports, per contract and model. (c) Market objective per world under continuous (grey) and batch (blue) trading; bars are means.

Limitations

Participants other than the news readers are stylized, and every parameter is set by us, so the numbers indicate the direction of mechanisms rather than their size in any market. The benefit of batch clearing depends on agents acting before slow traders in continuous trading. That the separate design’s exchange chooses continuous trading is an assumption about its objective. Six language-model worlds give large standard errors and should be read as a pilot. The news is generated from templates and is more regular than real reporting, and a model reading 30 contracts in one session may carry impressions between them.

References

  1. [1]Jacob Abernethy, Yiling Chen, and Jennifer Wortman Vaughan. Efficient Market Making via Convex Optimization, and a Connection to Online Learning. ACM Transactions on Economics and Computation, 1 (2):1–39, 2013. doi: https://doi.org/10.1145/2465769.2465777.
  2. [2]George A. Akerlof. The market for “lemons”: Quality uncertainty and the market mechanism. The Quarterly Journal of Economics, 84(3):488–500, 1970. doi: 10.2307/1879431.DOI
  3. [3]Guillermo Angeris and Tarun Chitra. Improved Price Oracles: Constant Function Market Makers. SSRN Electronic Journal, 2020.DOI
  4. [4]Kenneth J. Arrow, Robert Forsythe, Michael Gorham, Robert Hahn, Robin Hanson, John O. Ledyard, Saul Levmore, Robert Litan, Paul Milgrom, and Forrest D. Nelson. The Promise of Prediction Markets. Science, 2008. doi: 10.1126/science.1157679.DOI
  5. [5]Martino Banchio and Andrzej Skrzypacz. Artificial intelligence and auction design. In Proceedings of the 23rd ACM Conference on Economics and Computation, pages 30–31, 2022. doi: 10.1145/3490486.3538244.DOI
  6. [6]Rishi Bommasani, Kathleen A. Creel, Ananya Kumar, Dan Jurafsky, and Percy Liang. Picking on the Same Person: Does Algorithmic Monoculture Lead to Outcome Homogenization? In Advances in Neural Information Processing Systems 35, 2022. doi: 10.52202/068431-0265.DOI
  7. [7]Eric Budish, Peter Cramton, and John Shim. The high-frequency trading arms race: Frequent batch auctions as a market design response. The Quarterly Journal of Economics, 130(4):1547–1621, 2015. doi: 10.1093/qje/qjv027.
  8. [8]David Byrd, Maria Hybinette, and Tucker Hybinette Balch. ABIDES: Towards high-fidelity multi-agent market simulation. In Proceedings of the 2020 ACM SIGSIM Conference on Principles of Advanced Discrete Simulation, pages 11–22, 2020. doi: 10.1145/3384441.3395986.
  9. [9]Giulio Caldarelli. Understanding the Blockchain Oracle Problem: A Call for Action. Information, 11(11):509, 2020. doi: https://doi.org/10.3390/info11110509.
  10. [10]Emilio Calvano, Giacomo Calzolari, Vincenzo Denicolò, and Sergio Pastorello. Artificial intelligence, algorithmic pricing, and collusion. American Economic Review, 110(10):3267–3297, 2020. doi: 10.1257/aer.20190623.DOI
  11. [11]Yupeng Cao, Haohang Li, Weijin Liu, Wenbo Cao, Anke Xu, Lingfei Qian, Xueqing Peng, Minxue Tang, Zhiyuan Yao, Jimin Huang, et al. FinTrace: Holistic trajectory-level evaluation of LLM tool calling for long-horizon financial tasks, 2026.arxiv.org/abs/2604.10015
  12. [12]Álvaro Cartea, Patrick Chang, José Penalva, and Harrison Waldon. Algorithmic collusion and a folk theorem from learning with bounded rationality. Games and Economic Behavior, 157:1–21, 2026a. doi: 10.1016/j.geb.2025.11.012.
  13. [13]Álvaro Cartea, Jean-Edouard Colliard, Thierry Foucault, Peter Hoffmann, Rob Graumans, and Jean-David Sigaux. Artificial intelligence and financial markets. ECB Discussion Paper Series 32, European Central Bank, 2026b.DOI
  14. [14]Mithun Chakraborty and Sanmay Das. Trading on a rigged game: outcome manipulation in prediction markets. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, pages 158–164, 2016. URL https://www.ijcai.org/Proceedings/16/Papers/030.pdf.
  15. [15]Tatiana Chakravorti, Vaibhav Singh, Sarah Rajtmaier, Michael McLaughlin, Robert Fraleigh, Christopher Griffin, Anthony Kwasnica, David Pennock, and C. Lee Giles. Artificial prediction markets present a novel opportunity for human-AI collaboration. arXiv preprint arXiv:2211.16590, 2022. URL https://arxiv.org/abs/2211.16590.
  16. [16]Vincent Conitzer and Tuomas Sandholm. Complexity of mechanism design. In Proceedings of the 18th Conference on Uncertainty in Artificial Intelligence, 2002. doi: 10.48550/arXiv.cs/0205075.arxiv.org/abs/1408.1486
  17. [17]Philip Daian, Steven Goldfeder, Tyler Kell, Yunqi Li, Xueyuan Zhao, Iddo Bentov, Lorenz Breidenbach, and Ari Juels. Flash boys 2.0: Frontrunning in decentralized exchanges, miner extractable value, and consensus instability. In 2020 IEEE Symposium on Security and Privacy, pages 910–927, 2020. doi: 10.1109/SP40000.2020.00040.DOI
  18. [18]Paul Duetting, Zhe Feng, Harikrishna Narasimhan, David Parkes, and Sai Srivatsa Ravindranath. Optimal auctions through deep learning. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 1706–1715. PMLR, 2019. URL https://proceedings.mlr.press/v97/duetting19a.html.
  19. [19]Liran Einav, Chiara Farronato, Jonathan Levin, and Neel Sundaresan. Auctions versus posted prices in online markets. Journal of Political Economy, 126(1):178–215, 2018. doi: 10.1086/695529.
  20. [20]Duanyu Feng, Yongfu Dai, Jimin Huang, Yifang Zhang, Qianqian Xie, Weiguang Han, Zhengyu Chen, Alejandro Lopez-Lira, and Hao Wang. Empowering many, biasing a few: Generalist credit scoring through large language models, 2023.arxiv.org/abs/2310.00566
  21. [21]Ray Fisman and Tim Sullivan. The Inner Lives of Markets: How People Shape Them—And They Shape Us. PublicAffairs, New York, 2016. ISBN 9781610394925.
  22. [22]Rupert Freeman, Sébastien Lahaie, and David M. Pennock. Crowdsourced Outcome Determination in Prediction Markets. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 31, 2017. doi: https://doi.org/10.1609/aaai.v31i1.10594.
  23. [23]Anaïs Galdin and Jesse Silbert. Making talk cheap: Generative AI and labor market signaling. SSRN working paper, 2025.arxiv.org/abs/2511.08785
  24. [24]Polydoros Giannouris, Yuechen Jiang, Lingfei Qian, Yuyan Wang, Xueqing Peng, Jimin Huang, Guojun Xiong, and Sophia Ananiadou. Moira: Language-driven hierarchical reinforcement learning for pair trading, 2026.
  25. [25]Tilmann Gneiting and Adrian E. Raftery. Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association, 102(477):359–378, 2007. doi: https://doi.org/10.1198/016214506000001437.
  26. [26]Sanford J. Grossman and Joseph E. Stiglitz. On the impossibility of informationally efficient markets. American Economic Review, 70(3):393–408, 1980.DOI
  27. [27]Danny Halawi, Fred Zhang, Chen Yueh-Han, and Jacob Steinhardt. Approaching Human-Level Forecasting with Language Models. In Advances in Neural Information Processing Systems 37, 2024. doi: 10.52202/079017-1598.DOI
  28. [28]Yi Han, Yan Wang, Lingfei Qian, Haohang Li, Yupeng Cao, Yueru He, Xueqing Peng, Nanhan Shen, Yitao Xu, Yankai Chen, et al. Can LLM agents be CFOs? benchmarking long-horizon resource allocation in an uncertain enterprise environment, 2026.arxiv.org/abs/2603.23638
  29. [29]Robin Hanson. Logarithmic market scoring rules for modular combinatorial information aggregation. The Journal of Prediction Markets, 1(1):3–15, 2007. doi: 10.5750/jpm.v1i1.417.
  30. [30]Moritz Hardt, Nimrod Megiddo, Christos Papadimitriou, and Mary Wootters. Strategic classification. In Proceedings of the 2016 ACM Conference on Innovations in Theoretical Computer Science, pages 111–122, 2016. doi: 10.1145/2840728.2840730.DOI
  31. [31]Patrick Hemmer, Max Schemmer, Michael Vössing, and Niklas Kühl. Human-AI Complementarity in Hybrid Intelligence Systems: A Structured Literature Review. In PACIS 2021 Proceedings, page 78, 2021. URL https://aisel.aisnet.org/pacis2021/78/.
  32. [32]Jack Hirshleifer. The private and social value of information and the reward to inventive activity. The American Economic Review, 61(4):561–574, 1971.
  33. [33]Zoë Hitzig, Sylvie Carr, Tess Cotter, Kevin Troy, Kyle Turman, Maxim Massenkoff, and Peter McCrory. Project swap: What happens when agents trade for us? Anthropic research report, 2026. URL https://www.anthropic.com/research/project-swap.
  34. [34]Yuechen Jiang, Zhiwei Liu, Yupeng Cao, Yueru He, Ziyang Xu, Chen Xu, Zhiyang Deng, Prayag Tiwari, Xi Chen, Alejandro Lopez-Lira, et al. All that glisters is not gold: A benchmark for reference-free counterfactual financial misinformation detection, 2026.
  35. [35]Justin P. Johnson, Andrew Rhodes, and Matthijs Wildenbeest. Platform design when sellers use pricing algorithms. Econometrica, 91(5):1841–1879, 2023. doi: 10.3982/ECTA19978.
  36. [36]Jon Kleinberg and Manish Raghavan. Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences, 118(22), 2021. doi: 10.1073/pnas.2018340118.DOI
  37. [37]Haohang Li, Yupeng Cao, Yangyang Yu, Shashidhar Reddy Javaji, Zhiyang Deng, Yueru He, Yuechen Jiang, Zining Zhu, Koduvayur Subbalakshmi, Guojun Xiong, et al. INVESTORBENCH: A benchmark for financial decision-making tasks with LLM-based agent, 2024.arxiv.org/abs/2412.18174
  38. [38]Shengyuan Colin Lin, Felix Tian, Keyi Wang, Xingjian Zhao, Jimin Huang, Qiangqian Xie, Luca Borella, Matt White, Christina Dan Wang, Kairong Xiao, et al. Open FinLLM leaderboard: Towards financial AI readiness, 2025.arxiv.org/abs/2501.10963
  39. [39]Xiao-Yang Liu, Hongyang Yang, Qian Chen, Runjia Zhang, Liuqing Yang, Bowen Xiao, and Christina Dan Wang. FinRL: A Deep Reinforcement Learning Library for Automated Stock Trading in Quantitative Finance. arXiv preprint arXiv:2011.09607, 2020. URL https://arxiv.org/abs/2011.09607.
  40. [40]Zhiwei Liu, Xin Zhang, Kailai Yang, Qianqian Xie, Jimin Huang, and Sophia Ananiadou. FMDLlama: Financial misinformation detection based on large language models, 2024.DOI
  41. [41]Alejandro Lopez-Lira and Yuehua Tang. Can ChatGPT forecast stock price movements? return predictability and large language models. Journal of Financial Economics, 184:104335, 2026. doi: 10.1016/j.jfineco.2026.104335.
  42. [42]Robert Merton and Zvi Bodie. The Design of Financial Systems: Towards a Synthesis of Function and Structure. Technical Report 10620, National Bureau of Economic Research, 2004.
  43. [43]Robert C. Merton. A Functional Perspective of Financial Intermediation. Financial Management, 24(2):23, 1995. doi: https://doi.org/10.2307/3665532.
  44. [44]Nolan Miller, Paul Resnick, and Richard Zeckhauser. Eliciting informative feedback: The peer-prediction method. Management Science, 51(9):1359–1373, 2005. doi: 10.1287/mnsc.1050.0379.DOI
  45. [45]Michael Ostrovsky. Information Aggregation in Dynamic Markets With Strategic Traders. Econometrica, 80 (6):2595–2647, 2012. doi: https://doi.org/10.3982/ecta8479.
  46. [46]Parag A. Pathak and Tayfun Sönmez. Leveling the playing field: Sincere and sophisticated players in the Boston mechanism. American Economic Review, 98(4):1636–1652, 2008. doi: 10.1257/aer.98.4.1636.DOI
  47. [47]Xueqing Peng, Triantafillos Papadopoulos, Efstathia Soufleri, Polydoros Giannouris, Ruoyu Xiang, Yan Wang, Lingfei Qian, Jimin Huang, Qianqian Xie, and Sophia Ananiadou. Plutus: Benchmarking large language models in low-resource Greek finance, 2025a.arxiv.org/abs/2502.18772
  48. [48]Xueqing Peng, Lingfei Qian, Yan Wang, Ruoyu Xiang, Yueru He, Yang Ren, Mingyang Jiang, Vincent Jim Zhang, Yuqing Guo, Jeff Zhao, et al. MultiFinBen: Benchmarking large language models for multilingual and multimodal financial application, 2025b.
  49. [49]Xueqing Peng, Zhuohan Xie, Yupeng Cao, Haohang Li, Lingfei Qian, Yan Wang, Vincent Jim Zhang, Huan He, Xuguang Ai, Linhai Ma, et al. Herculean: An agentic benchmark for financial intelligence, 2026.arxiv.org/abs/2605.14355
  50. [50]Juan C. Perdomo, Tijana Zrnic, Celestine Mendler-Dünner, and Moritz Hardt. Performative prediction. In Proceedings of the 37th International Conference on Machine Learning, 2020. doi: 10.48550/arXiv.2002.06673.DOI
  51. [51]Dražen Prelec. A Bayesian truth serum for subjective data. Science, 306(5695):462–466, 2004. doi: 10.1126/science.1102081.DOI
  52. [52]Canice Prendergast. How food banks use markets to feed the poor. Journal of Economic Perspectives, 31(4):145–162, 2017. doi: 10.1257/jep.31.4.145.DOI
  53. [53]Lingfei Qian, Xueqing Peng, Yan Wang, Vincent Jim Zhang, Huan He, Hanley Smith, Yi Han, Yueru He, Haohang Li, Yupeng Cao, et al. When agents trade: Live multi-market trading benchmark for LLM agents, 2025.arxiv.org/abs/2510.11695
  54. [54]Matthew Riemer, Tommaso Tosato, Amin Memarian, Maximilian Puelma Touzel, Glen Berseth, Irina Rish, and Guillaume Dumas. Position: Collusion Risks Among AI Reasoning Agents Justify Certification Requirements for Making Market Decisions, 2026.arxiv.org/abs/2608.18078
  55. [55]Lee Ross. The intuitive psychologist and his shortcomings: Distortions in the attribution process. In Advances in Experimental Social Psychology, volume 10, pages 173–220. Academic Press, 1977. doi: 10.1016/S0065-2601(08)60357-3.DOI
  56. [56]Alvin E. Roth. The Economist as Engineer: Game Theory, Experimentation, and Computation as Tools for Design Economics. Econometrica, 2002. doi: 10.1111/1468-0262.00335.DOI
  57. [57]Philipp Schoenegger, Indre Tuminauskaite, Peter S. Park, Rafael Valdece Sousa Bastos, and Philip E. Tetlock. Wisdom of the silicon crowd: LLM ensemble prediction capabilities rival human crowd accuracy. Science Advances, 2024. doi: 10.1126/sciadv.adp1528.DOI
  58. [58]Philipp Schoenegger, Peter S. Park, Ezra Karger, Sean Trott, and Philip E. Tetlock. AI-Augmented Predictions: LLM Assistants Improve Human Forecasting Accuracy. ACM Transactions on Interactive Intelligent Systems, 15(1):1–25, 2025. doi: https://doi.org/10.1145/3707649.
  59. [59]Joar Skalse, Nikolaus Howe, Dmitrii Krasheninnikov, and David Krueger. Defining and Characterizing Reward Hacking, 2022.arxiv.org/abs/2209.13085
  60. [60]Thomas Spooner, John Fearnley, Rahul Savani, and Andreas Koukorinis. Market Making via Reinforcement Learning. arXiv preprint arXiv:1804.04216, 2018. URL https://arxiv.org/abs/1804.04216.
  61. [61]Kevin K. Troy, Dylan Shields, Keir Bradwell, and Peter McCrory. Project deal: Our Claude-run marketplace experiment. Anthropic research report, 2026. URL https://www.anthropic.com/features/project-deal.
  62. [62]Michelle Vaccaro, Abdullah Almaatouq, and Thomas W. Malone. When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8(12):2293–2303, 2024. doi: https://doi.org/10.1038/s41562-024-02024-1.
  63. [63]Yan Wang, Keyi Wang, Shanshan Yang, Jaisal Patel, Jeff Zhao, Fengran Mo, Xueqing Peng, Lingfei Qian, Yankai Chen, Víctor Gutiérrez-Basulto, et al. FinAuditing: A financial taxonomy-structured multi-document benchmark for evaluating LLMs, 2025.arxiv.org/abs/2510.08886
  64. [64]Yan Wang, Yi Han, Lingfei Qian, Yueru He, Xueqing Peng, Dongji Feng, Zhuohan Xie, Vincent Jim Zhang, Rosie Guo, Fengran Mo, et al. Conv-FinRe: A conversational and longitudinal benchmark for utility-grounded financial recommendation, 2026.arxiv.org/abs/2602.16990
  65. [65]Justin Wolfers and Eric Zitzewitz. Prediction Markets. Journal of Economic Perspectives, 2004. doi: 10.1257/0895330041371321.DOI
  66. [66]Mengxi Xiao, Zihao Jiang, Lingfei Qian, Zhengyu Chen, Yueru He, Yijing Xu, Yuecheng Jiang, Dong Li, Ruey-Ling Weng, Min Peng, et al. Retrieval-augmented large language models for financial time series forecasting, 2025.arxiv.org/abs/2502.05878
  67. [67]Yijia Xiao, Edward Sun, Di Luo, and Wei Wang. TradingAgents: Multi-agents LLM financial trading framework. arXiv preprint arXiv:2412.20138, 2024. doi: 10.48550/arXiv.2412.20138.
  68. [68]Qianqian Xie, Weiguang Han, Xiao Zhang, Yanzhao Lai, Min Peng, Alejandro Lopez-Lira, and Jimin Huang. PIXIU: A Large Language Model, Instruction Data and Evaluation Benchmark for Finance, 2023.arxiv.org/abs/2306.05443
  69. [69]Qianqian Xie, Weiguang Han, Zhengyu Chen, Ruoyu Xiang, Xiao Zhang, Yueru He, Mengxi Xiao, Dong Li, Yongfu Dai, and Duanyu Feng. FinBen: A Holistic Financial Benchmark for Large Language Models. In Advances in Neural Information Processing Systems 37, 2024. doi: 10.52202/079017-3033.DOI
  70. [70]Zhuohan Xie, Daniil Orel, Rushil Thareja, Dhruv Sahnan, Hachem Madmoun, Fan Zhang, Debopriyo Banerjee, Georgi Georgiev, Xueqing Peng, Lingfei Qian, et al. FinChain: A symbolic benchmark for verifiable chain-of-thought financial reasoning, 2025.
  71. [71]Guojun Xiong, Zhiyang Deng, Keyi Wang, Yupeng Cao, Haohang Li, Yangyang Yu, Xueqing Peng, Mingquan Lin, Kaleb E Smith, Xiao-Yang Liu, et al. FLAG-Trader: Fusion LLM-Agent with gradient-based reinforcement learning for financial trading, 2025.arxiv.org/abs/2502.11433
  72. [72]Yangyang Yu, Haohang Li, Zhi Chen, Yuechen Jiang, Yang Li, Denghui Zhang, Rong Liu, Jordan W. Suchow, and Khaldoun Khashanah. FinMem: A performance-enhanced LLM trading agent with layered memory and character design. In Proceedings of the AAAI Symposium Series, volume 3, pages 595–597, 2024a. doi: 10.1609/aaaiss.v3i1.31290.DOI
  73. [73]Yangyang Yu, Zhiyuan Yao, Haohang Li, Zhiyang Deng, Yupeng Cao, Zhi Chen, Jordan W. Suchow, Rong Liu, Zhenyu Cui, Zhaozhuo Xu, et al. FinCon: A synthesized LLM multi-agent system with conceptual verbal reinforcement for enhanced financial decision making, 2024b.arxiv.org/abs/2407.06567
  74. [74]Wentao Zhang, Lingxuan Zhao, Haochong Xia, Shuo Sun, Jiaze Sun, Molei Qin, Xinyi Li, Yuqing Zhao, Yilei Zhao, Xinyu Cai, Longtao Zheng, Xinrun Wang, and Bo An. A multimodal foundation agent for financial trading: Tool-augmented, diversified, and general. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024a. doi: 10.1145/3637528.3671801.DOI
  75. [75]Stephan Zheng, Alexander Trott, Sunil Srinivasa, David C. Parkes, and Richard Socher. The AI economist: Taxation policy design via two-level deep multiagent reinforcement learning. Science Advances, 8(18): eabk2607, 2022. doi: 10.1126/sciadv.abk2607.DOI

Paper details

vlnc:W_tf7dbmx2
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Codex, Anthropic Claude

View statement
Looking for related papers…
Contents