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

Yankai Chen, Xueqing Peng, Alejandro Lopez-Lira, Yuehua Tang, Michał Dzieliński, Yijia Zhao, Jimin Huang, Xue Liu, Sophia AnaniadouarXiv 2026 · arXiv

TL;DR

This paper argues that we should stop treating financial markets as fixed environments for AI agents and instead design the market itself—its rules, liquidity, and settlement—jointly with the agents. By treating market functions as learnable components (HyperMarkets), the authors demonstrate a 79% improvement in market-level objectives, showing that how a market is built determines whether AI helps or exploits its participants.

What Problem Does This Paper Address?

In traditional machine learning, we usually train an agent to perform well in a static environment. In finance, this means building "smarter" traders to operate in existing markets. However, a market isn't a static backdrop; it’s a system of rules that defines the "reward" for every participant.

The authors point out a critical flaw: if you place a highly capable AI agent into a market designed for humans (who are slow and have limited attention), the agent often learns to exploit the market's mechanical weaknesses (like millisecond-level arbitrage) rather than producing useful information. The paper addresses the need for HyperMarkets: markets where liquidity, information production, and trading rules are designed jointly to achieve collective intelligence rather than just individual profit.

Figure 1: The shift from agent-centered research to HyperMarket research, where the market functions themselves are designed toward a collective objective.

Methods

The authors propose a framework where four core market functions are treated as learnable or adaptable problems:

  1. Liquidity Provision: Moving from human attention to sequential decision-making under inventory risk.
  2. Information Production: Rewarding incremental, independent evidence rather than just a high volume of text.
  3. Outcome Determination: Using automated evidence assessment with human-in-the-loop adjudication.
  4. Trading Rules: Moving from continuous trading (which rewards speed) to mechanisms like batch auctions (which reward information).

They formalize this as a bilevel optimization problem: the "outer" designer chooses market rules ($\theta$) to maximize a market-level objective (like price accuracy), while the "inner" agents ($\pi^*$) learn to maximize their own rewards given those rules.

To test this, they ran a pre-registered simulation using Claude 3.5 Sonnet (strong agent) and Claude 3.5 Haiku (weak agent) acting as traders for human principals, alongside "slow" human-like traders.

Datasets

The paper evaluates the current landscape of financial AI using several existing benchmarks:

  • Static Tasks: PIXIU, FinBen, MultiFinBen.
  • Agentic Workflows: FinTrace, CFO arena, InvestorBench.
  • Live/Simulated Markets: ABIDES, FinRL, Agent Market Arena.

For their own simulation, they generated a custom dataset of 60 candidate contracts (e.g., "Will dataset X be released by day 12?") and a corpus of 481 independent news items and 992 re-reports (spam/echoes) to test how agents handle information quality.

Figure 4: A timeline showing how financial AI evaluation has moved from static tasks to live markets, but has consistently held the market rules fixed.

Key Insights and Ideas

  • Liquidity is the Bottleneck: In simulations, making liquidity "ideal" first provided the biggest boost to the market objective. Information only becomes valuable once the market is liquid enough to absorb it.
  • You Get What You Pay For: If the market monitor is paid "per report," the system is flooded with redundant "echo" reports that lead to mis-settled contracts. Paying only for new, independent evidence is required to make a stronger agent (like Sonnet) actually provide more value to its principal than a weaker one.
  • Batching Protects Humans: Continuous trading rewards the speed of AI agents. Switching to uniform-price batch clearing (grouping trades into discrete time intervals) protects "slow" traders and brings prices closer to the true information value.
  • Joint Design Wins: Tuning all market functions together is significantly more effective than tuning each one in isolation. The interactions between how you provide liquidity and how you settle claims are too complex to optimize separately.
Figure 5: The "Liquidity First" insight—improving liquidity provides the greatest initial gain to the market objective $J$.

Key Results

  • 79% Objective Increase: Jointly designing market functions raised the market-level objective (a mix of price accuracy and fair settlement) by 79% over separately tuned functions.
  • Spam Reduction: Paying for new evidence instead of per-report cut mis-settled contracts from 2.23 to 0.80 per world.
  • Profit Redistribution: In mixed markets, batch clearing turned slow traders' average results from a loss of -3.4 to a profit of +9.2, effectively moving profit from fast AI arbitrageurs back to the participants they serve.
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.

Limitations and Future Work

  • Stylized Simulation: The participants (other than the LLM news readers) were stylized agents with fixed behaviors. Real human behavior in these markets remains untested.
  • Small Sample Size: The LLM-based "worlds" were limited to a pilot of six, leading to large standard errors in some metrics.
  • Template-Based News: The news items used in the simulation were generated from templates, which is more regular and predictable than real-world financial reporting.
  • Future Direction: The authors call for benchmark builders to "expose the market" as a variable, allowing researchers to compete not just on who has the best trader, but on who can design the most efficient and fair market mechanism.

Summary updated

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.
BibTeX
@misc{chen2026market,
  author = {Yankai Chen and Xueqing Peng and Alejandro Lopez-Lira and Yuehua Tang and Michał Dzieliński and Yijia Zhao and Jimin Huang and Xue Liu and Sophia Ananiadou},
  title = {{The Market Is the Reward: Design Markets, Not Just Traders for Human-AI Collective Intelligence}},
  year = {2026},
  eprint = {2611.14564},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.14564},
}
Figure