Decitron: World's First General-Purpose Decision-Making AI Model Released

Generating answers is becoming a fundamental capability of AI, but reasoning about the real world is the more challenging problem for the next stage.

Real decision-making is not a static Q&A. It occurs in a dynamic world where multiple actors act simultaneously, information is incomplete, and uncertainty prevails.

For example, if a newly open-sourced model suddenly slashes costs, will leading model companies follow suit with price cuts? How will cloud providers, developers, and application companies realign themselves? In such decisions, one party’s action alters the conditions facing the others, and each new response continues to reshape the landscape. It’s like a game where all players move simultaneously—each move changes the board.

Against this backdrop, after the release of Zhongke Wenge’s Decitron (abbreviated as Decitron), it was hailed by multiple leading tech media as the "world’s first general-purpose decision-making large model." This designation does not emphasize the first appearance of any single technology—such as world models, multi-agent systems, or game-solving—but rather highlights that Decitron is the first to unify these capabilities within a single decision-making framework designed for open worlds, creating a complete closed loop of continuous modeling, simulation, solving, and calibration.

The term “general” refers to its attempt to abstract complex decision-making problems across different domains into a common structure—world state, participating agents, action space, constraint relationships, multi-round interactions, and uncertain outcomes—and to continuously reason and solve within the same framework.

Technical report URL: https://chinaxiv.org/abs/202608.00064

On August 3, the Decitron decision engine released its comprehensive technical report, publicly outlining its three core technological contributions for the first time:

First, we introduce MetaWorld, an explicit world model that organizes multi-source information into a structured, persistent, and computable world state with causal modeling; second, we employ a State–Action–Outcome (SAO) formalism to transform complex decisions into computable expressions and reasoning formats; third, we construct AutoABM, a hybrid reasoning architecture that integrates multi-agent simulation, game-theoretic reasoning, and formal optimization to solve complex decision problems such as game equilibrium, strategy planning, and constraint satisfaction through a multi-agent decision inference system. Together, these three components form the technical foundation of the decision engine, aiming to transform AI systems’ one-time judgments into a continuous process of computation and inference that evolves with the changing world.

Currently, several key capabilities required by AI for "decision simulation"—such as world models, multi-agent systems, game-theoretic reasoning, and probabilistic forecasting—have already emerged one by one. The real challenge lies in ensuring that when AI transitions from closed tasks to the open real world, these capabilities must work together cohesively around a single, continuously updated world state, rather than providing isolated analyses within a single prompt.

First, consider world models. When people hear "World Model," they often think of physical world models like Sora, World Labs, and V-JEPA, which primarily focus on visual, spatial, physical, or robotic environments. In contrast, MetaWorld, introduced by Decision Machine, addresses far more complex decision-making scenarios: open systems such as economics, policy, corporate competition, and geopolitics, involving a much more intricate set of agents, variables, relationships, and constraints.

MetaWorld can be seen as an extension of world models into social systems, requiring the handling of three challenging issues: rule disparities, reflexivity (bidirectional interaction and circular influences between cause and effect), and irreducible uncertainty. This also raises a direct question: if the assessment of the current world is biased, will subsequent projections also become skewed?

Once integrated into a social system, maintaining environmental states alone is insufficient; AI must also assess how other participants will act. Projects like CICERO from Meta and Peking University have already demonstrated the ability of multiple agents to interact and engage in strategic博弈 under limited information. For real-world decision-making in open-ended environments, after simulating the actions of all parties, it is further necessary to evaluate whether strategies are viable, identify potential equilibria in the博弈, and determine the likelihoods and conditions of different future outcomes.

The significance of the decision engine lies herein: its hybrid reasoning architecture, AutoABM, unifies LLM-based semantic understanding, SAO structured representation, multi-agent modeling and simulation, and strategic reasoning and optimization onto a single implementation pipeline, seamlessly connecting the four steps of “understanding the world—simulating scenarios—deriving decisions—evaluating future outcomes.”

The decision-making engine consists of six layers—input and task, data and knowledge, modeling and formalization, simulation and reasoning, prediction and calibration, reporting and explanation—forming a complete pipeline from real-world information to decision outputs.

Next, we use the example of “How might a complex trade conflict evolve over the next six months?” to show how the decision engine starts from real-world information, gradually builds a model of the current state of the world, simulates the actions of all parties, explores potential future pathways, and identifies the conditions under which each pathway may occur.

Step one: Don’t rush to predict the future—first, understand what is happening now and which facts are trustworthy.

Real-world information is often noisy, potentially redundant, conflicting, or outdated. The decision engine first transforms the user’s question into a structured task, clearly defining what needs to be answered, which entities are involved, the time period to observe, and any constraints. In this case, the system first identifies the key parties involved in the conflict, the relevant time frame, and the current points of contention; it then organizes conditions that could influence the situation, such as tariffs, supply chains, industrial policies, and corporate behavior. Following this, the system gathers information from diverse sources—including news articles, social media, and research reports—around these key elements.

Figure 2: Data Processing Flowchart

The collected information is first deduplicated, clustered, and processed for event extraction, organized into an event timeline and a set of finer-grained evidence units. The system then evaluates the quality of each piece of evidence using an Evidence Quality Score (EQS), assessing factors such as relevance to the current query, source credibility, timeliness and completeness of content, and consistency across different sources. Based on the confidence level of each piece of evidence, it is routed accordingly—high-confidence evidence proceeds directly to subsequent reasoning steps.

If the existing evidence is insufficient or conflicting sources arise, the system continues executing ReAct-based alternating reasoning and retrieval, adjusting the next search based on currently missing information and incorporating newly found data into the existing timeline.

Figure 3: Using a ReAct-based alternating mechanism of reasoning and retrieval

Step two: Convert the filtered facts into a continuously updated "world state." The facts tell the system what has occurred; next, determine how these changes, when combined, define the current state of the world.

The decision engine formalizes this process using a State–Action–Outcome (SAO) representation, where State describes the current state of the world, Action records the actions taken by each agent under that state, and Outcome captures the resulting impacts such as gains, losses, costs, and risks. Notably, State distinguishes between macro-level states and more granular, quantifiable states. In this case, escalating friction may serve as a macro-level state, while tariff levels, commodity prices, and the extent of supply chain disruption can be included in the finer-grained state records.

In addition, the system separately accounts for external shocks, such as sudden policy changes, environmental shifts, or other events beyond the Agent’s control. After an action and an external event occur, the world state is updated, and this updated state becomes the starting point for the next action. This cycle repeats continuously, forming a continuous SAO trajectory.

Figure 4: SAO Process Flow

MetaWorld is responsible for maintaining these states, outputting a structured world state that can be computed and updated. The state can be divided into three layers: the Environment Layer records the shared overall environment for all participants, such as phase of the situation and resource prices; the Agent Metrics Layer tracks each agent’s resources, capabilities, and progress toward goals; and the Relationship Matrix Layer captures changes in relationships between different agents.

MetaWorld also employs a Causal Directed Acyclic Graph (Causal DAG) to constrain how states change: potential influences between variables are organized into a causal graph during the modeling phase. During inference, the system updates node states and the weights of causal edges, but does not regenerate the causal graph each round.

This is crucial for long-term simulations. Previous tariff changes affected prices, which in turn altered corporate decisions and the responses of other participants—these changes must be preserved to continue the simulation. MetaWorld serves as this continuously updated “ledger of the world.”

Step three: With the world state established, the various agents begin to act. Traditional Agent-Based Modeling (ABM) typically requires experts to predefine participants, behavioral rules, environmental variables, and interaction mechanisms. Changing the question often means rebuilding the entire model. The Decision Engine’s AutoABM automates this step: given a natural language question, the system integrates external evidence to extract participants, goals and constraints, action spaces, environmental variables, and causal relationships, then constructs a multi-agent simulation environment accordingly.

A trade conflict may involve multiple agents—such as governments, corporations, industry associations, and consumers—each with their own objectives, resources, risk preferences, and non-negotiable boundaries. These agents lack a god’s-eye view and can only act based on the partial information they possess.

In each round of action, the Agent first retrieves previous interactions and historical trajectories, then assesses the opportunities, risks, and constraints in the current situation, generates several optional strategies, and makes a selection. Moreover, each round does not focus on a single direction; instead, the system simultaneously explores multiple distinct world states, including relatively stable equilibrium paths, more optimistic or pessimistic evolutionary trajectories, and low-probability but high-impact branches that would significantly alter the situation if realized. The Agent then filters and validates these paths, updates the world state, and proceeds to the next round. As all parties continue to act, some paths lose their validity while others become more likely.

Figure 5: Agent Behavior Paradigm

Beyond multi-agent simulations, complex decision-making requires a "mathematical referee." While multi-agent systems can generate many seemingly plausible future pathways, merely envisioning them does not mean they are achievable—real-world decisions are constrained by hard limits such as budget, resources, time, policy boundaries, and balanced interests among stakeholders.

The decision engine integrates a formal solving capability covering five categories of problems: classical game theory, resource allocation and optimization, path planning and scheduling, probabilistic inference under uncertainty, and constraint satisfaction. For example, it can compute how to allocate resources within a limited budget, determine potential equilibria in multi-party games, or assess whether a strategy breaches predefined thresholds. In game theory, the system includes nine classic atomic games—such as the Prisoner’s Dilemma, Stag Hunt, Chicken Game, and Zero-Sum Games—covering 144 Robinson–Goforth 2×2 game topologies to identify and solve diverse game structures.

The focus of this layer is on feeding the candidate strategies generated by multi-agent systems back into computable frameworks of games, constraints, and optimizations, thereby transforming "semantically reasonable" into "formally verifiable."

On the TMGBench benchmark for testing multi-agent game capabilities, the decision engine achieves a equilibrium accuracy of 99.4%, a topology identification accuracy of 100%, an average solving time of 0.8 seconds, and an interpretability score of 4.3/5.0.

Table 2: Five Main Types of Solvable Problems

The division of labor between multi-agent simulation and formal solving is clear: the former expands possible pathways, while the latter verifies whether these pathways hold under constraints and game structures.

At the final stage of the scenario analysis, one key question remains: What is the likelihood of each of the previously outlined future pathways? To address this, the decision engine incorporates a forecasting and calibration process.

Building on the candidate paths identified earlier, this layer further addresses uncertainty: the system combines model inference results with external calibration signals to assign probabilities to different outcomes, then calibrates these probabilities using prediction market data, historical forecasting performance, and probability scoring rules. The decision engine no longer merely outputs a list of “what might happen” scenarios—each path now comes with an associated probability judgment, continuously updated as new evidence and inference results emerge.

We can think of it as a dynamic, evolving "map of the future": ongoing escalation of trade tensions, temporary de-escalation between parties, or new unforeseen variables could all become different branches. The system evaluates the likelihood of each path and identifies which changes could make one path more probable than others.

Figure 6: Probability Prediction Module

Thus, the decision-making engine has connected a complete cycle of decision simulation.

The report used several sets of experiments to test the decision engine’s inference system, with the most noteworthy results being those on "event inference and probability prediction."

First, examine the self-built event prediction set to observe whether structured reasoning improves event forecasting. The team constructed a dataset containing 74 events and 370 binary judgment questions, covering multiple high-uncertainty event categories and incorporating varying prediction horizons. Under the team’s evaluation setup, the decision engine achieved an overall accuracy of 78.41%.

Additionally, in the LLM-NEWS settings, for the short-term group (3–30 days), the decision engine achieved a result of 72.80%, while GPT-5.5 and Claude-4.7 scored 35.43% and 44.62% respectively; as we move into medium- and long-term periods, this gap gradually narrows. This indicates that structured reasoning demonstrates a relatively greater advantage in short-term scenarios characterized by rapid changes and where historical patterns are less applicable.

Then, supplementary experiments on the public prediction market benchmark PolyBench.

The decision engine achieves an FFA (Final Forecast Accuracy) of 81.20%, an EVPA (Expected Volatility Prediction Accuracy) of 73.40%, and an MSE (Mean Squared Error) of 0.822, all of which outperform Gemini-3-Flash, MiMo-V2-Flash, and Grok-4.1-Fast. This indicates that the decision engine demonstrates superior performance in capturing the direction of market probability changes and controlling probability error.

Generative AI uses tokens as its fundamental unit of computation, addressing the question: "How should the next segment of content be generated?" Decision AI, by contrast, treats changes in the state of the world as its fundamental unit of computation, addressing the question: "How will a single action alter the future state of the world?" The former focuses on whether an answer is reasonable; the latter must also handle causality, real-world constraints, opponent responses, and shifting probabilities. These are not merely incremental improvements in capability, but fundamental shifts in computational paradigm.

Decision AI cannot naturally emerge merely through larger parameter scales and stronger language capabilities. Once entering the open world, foundational models remain important, but they shift from being the entire system to becoming components of a decision-making system. The unit of competition in AI also transitions from individual models to complete systems.

The technical significance of the decision engine lies in its ability to organize previously fragmented capabilities into a unified decision-making structure. The term “universal” does not imply predicting the future across all domains, but rather that problems from different domains can all be translated into states, actions, outcomes, constraints, and uncertainties, and then processed within the same reasoning and solution framework.

From this perspective, the true value of the "world's first general-purpose decision large model" lies not merely in claiming the title of "first," but in establishing a relatively complete technical framework for general-purpose decision large models: explicitly describing the world through states, executing actions through multi-agent systems, testing strategies through game theory and optimization, and calibrating probabilities across different future pathways. Correspondingly, the evaluation standards for AI are also shifting—not only assessing whether answers are accurate, but also whether states remain consistent, whether reasoning can be verified, whether probabilities are reliable, and whether the system can promptly update in response to new information.

After all, an answer that sounds right doesn't equate to a decision that can withstand real-world changes.

This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), authored by someone focused on decision AI.