A team of five researchers proposes treating certain human decisions as the outcome of collective dynamics, rather than as the simple sum of individual choices. The paper, uploaded to arXiv on September 7, brings this idea into the debate on artificial general intelligence: according to the authors, future AGI systems could complement neural networks with modules capable of modeling mutual adaptation, sudden strategy shifts, and coordinated behavior.
The title of the paper, Adaptive Entangled Game Modules in Artificial General Intelligence, evokes concepts drawn from game theory, mathematical physics, and cognitive science. Yet the empirical test chosen is neither a video game nor an AI benchmark: it is intraday trading on the Chinese stock market. There, Haochen Li, Xinshuai Guo, Jingdong Ouyang, Wei Zhang, and Leilei Shi argue, traders' decision-making patterns offer observable traces of interactions more complex than the classical economic model of the independent, rational investor.
The scope of the story should be clarified right away. The paper is a preprint: as of its publication date on arXiv, no peer review is indicated in the record. The percentages and conclusions regarding AGI architecture are therefore claims made by the authors, not established evidence from the scientific community. The proposal is intriguing precisely because it is ambitious, but it bridges theoretical domains that are far apart: market data, human behavior, brain mechanisms, and the design of intelligent machines.
From the isolated trader to emergent group behavior
At the core of the work is a so-called non-local probability equation of “generalized behavioral intelligence,” abbreviated as GBI. In less formal terms, it is a mathematical framework that attempts to represent adaptive agents whose behaviors do not remain isolated: one agent's decisions can depend on the context and the choices of others in a distributed manner throughout the system. The authors describe the model's possible outcomes as eigenstates or modes—recurrent configurations that can be tested against empirical data.
The word “entangled” should be read with caution. The paper relies on language and tools inspired by probability waves and even cites quantum physics in its arXiv classification, but it does not prove that traders are physically linked by quantum phenomena. Its use refers to a dependency modeled as non-local between behaviors: a formal way of describing decisions that do not appear reducible to completely independent agents.
This approach contrasts with neoclassical finance, where the baseline assumption tends to favor rational, autonomous, and informed individuals. Real markets, however, are an environment where signals, imitation, news, and simultaneous reactions often make independence a narrow approximation. The authors' contribution does not lie in noting this gap—widely discussed in behavioral finance as well—but rather in seeking to trace it back to a single mathematical framework and then turn it into a component for artificial general intelligence.
What the Chinese data indicates according to the paper
In their empirical analysis, the authors state that adaptive and interdependent game modes account for 82% to 94% of observed decision-making patterns, with an overall figure of 89%. Modes attributed to purely independent agents, by contrast, reportedly remain below 5%. A share between 2% and 12% of behaviors is linked to adapting to news, events, and environmental changes occurring during the trading session.
The paper associates this latter component with two equilibrium states and sharp shifts in the reference point agents use to evaluate gains, losses, or market conditions. This is a relevant factor for simulation designers: a system that fails to update its reference point when the context changes can produce linear responses at the very moments people are rapidly altering their strategy.
However, the available abstract does not reveal all the details essential for assessing the scope and robustness of the findings: precise sample composition, time period covered, out-of-sample validation procedures, and comparisons with alternative models. Even a high proportion of variance or explained patterns does not automatically prove the validity of a theory regarding the underlying cognitive mechanisms. These checks require a comprehensive review of the methods, independent replications, and a transparent comparison with existing econometric and behavioral approaches.
The leap from observation to AGI
The most forward-looking part of the work focuses on the architecture of intelligent systems. The authors argue that artificial neural networks, while effective across numerous tasks, rely on massive amounts of parameters that are difficult to interpret. Rather than replacing them, their proposal is to combine them with simulations based on probability waves and adaptive game modules. The goal would be to enrich foundation models with an explicit representation of agent interactions and contextual shifts.
This vision introduces HPUs, human-like processing units: processing units designed to incorporate brain-inspired mechanisms and produce more compact, efficient, and robust AGI systems. The envisioned scenarios include embodied intelligence and robotics—fields where a machine must operate in shared spaces alongside people and other devices, interpreting expectations, cooperation, conflict, and incomplete signals.
For gaming, the most practical takeaway does not involve a promise of “AGI in video games,” but rather the theoretical possibility of building less rigidly scripted non-player agents. In principle, an NPC or a system-controlled opponent could adapt its strategy to the group's moves, shift priorities in the face of an unexpected event, and display forms of coordination without requiring hardcoded rules for every scenario. This is a design hypothesis, not a finding proven by the paper: the study presents neither a playable prototype nor benchmarks conducted within a game engine.
A practical question also remains open. A more explicit mathematical model may be more interpretable than a neural network with billions of parameters, but interpretability also depends on how its variables are estimated, data quality, and how closely the assumptions genuinely reflect the phenomenon. The architectural compactness cited by the authors must be measured on real-world tasks and hardware. The same holds true for reliability in robotics, where adaptive behavior is desirable only if it remains predictable and subject to safety constraints.
The neurological hypothesis and the missing test
The paper links its findings to the Liu-Chen-Ao hypothesis, described as a hypothesis of non-local and “entangled” nerve fibers in the brain. The authors argue that the collective behavior of traders could provide an indirect method to examine it, as observable decisions might reflect internal mechanisms of intelligence and behavioral psychology.
This is the document's most delicate claim. Trading data can show collective correlations and adaptations, but bridging the gap from such observations to a neurobiological explanation requires a much broader chain of evidence, including independent experiments and measurements directly tied to brain activity. The study frames this connection as empirical support for the hypothesis; for now, however, it is more accurate to view it as a research avenue suggested by the authors.
The next useful step will therefore be less narrative and more experimental: making GBI methods verifiable across different datasets, benchmarking them against competing models, and demonstrating whether the described modules yield measurable advantages in software or robotic agents. If this path holds up, the concept of training an AI not only on regularities in data but also on the dynamics between subjects could find applications in simulations, gaming worlds, and multi-agent systems. If it does not hold up, the preprint will still remain an original attempt to bring game theory, markets, and artificial cognitive architectures into the same framework.
