China's Tezign Technology Releases Subjective World Model, Achieving 85% Accuracy in Behavior Simulation

en.Wedoany.com Reported - China's Tezign Technology released the Subjective World Model at the 2026 World Artificial Intelligence Conference (WAIC), simulating real consumer decision-making behavior through four-layer heterogeneous data collaborative training. The model's design is based on the known fact of systematic deviation between "self-reported preferences" and "actual decision-making weights" in cognitive psychology, aiming to address the inherent limitations of large language models in modeling human decision-making.

The exhibition attracted over 1,100 enterprises, covering an area of 100,000 square meters. Tezign Technology's Subjective World Model (SWM) architecture comprises four layers. The first layer is the Expression Layer, which uses billions of native social platform corpora to construct high-dimensional mappings of user language style vectors with demographic and psychological characteristics. The second layer is the Narrative Layer, which models temporal causal graphs of behavioral motivation based on tens of thousands of hours of one-on-one in-depth interview corpora (each interview lasting 1 to 2 hours, generating 5,000 to 20,000 words of unstructured text); this layer emphasizes that real interview corpora possess cross-context causal consistency, which synthetic data cannot replace.

The third layer is the Cognitive Layer, which models individual value weight vectors and risk preference coefficients using standardized scales such as behavioral judgment questionnaires, the Schwartz Value Survey, and the Big Five personality inventory, aiming to quantify the gap between consumers' stated preferences and actual decision-making trade-offs. The fourth layer is the Behavioral Layer, which uses experimental data from economic games such as the Ultimatum Game, Public Goods Game, and Trust Game, along with real transaction records, for parametric individual modeling, outputting three behavioral economics parameters: loss aversion coefficient λ, hyperbolic time discount rate δ, and social norm sensitivity γ.

During inference, SWM maintains four types of joint states for each target persona: Expression Layer language style embeddings, Narrative Layer motivation causal graphs, Cognitive Layer value weight vectors, and Behavioral Layer economic parameters. The four layers jointly score to generate an AI Persona with intrinsic states. Unlike stateless large language models that generate responses independently each time, SWM's Persona can maintain consistent intrinsic states under continuous questioning. Quantitative results show a behavior simulation accuracy of 85% (benchmarked against in-depth human interviews), capable of generating over 300,000 AI Personas through social data diffusion, including more than 10,000 high-precision Personas directly trained from interview corpora, with a delivery time of less than 30 minutes.

Beyond the understanding side, Tezign Technology has developed the GEA (Generative Enterprise Agent) enterprise-level agent architecture for executing consumer-facing tasks. The orchestration layer of this architecture is driven by China's first registered divergent reasoning domain large model, the Creative Reasoning Model. The execution layer can invoke over 400 modular Agent Skills and coordinate more than 30 foundation models, with the Context System serving as the enterprise knowledge persistence storage layer. Currently, this architecture has been deployed at scale in business scenarios such as content marketing, insight research, product innovation, and design creation, with a monthly average Token deployment exceeding 10 billion, covering over 50 countries and regions worldwide.

Tezign stated that SWM's competitive barrier lies not in the replicability of its architectural design, but in the real interview corpora accumulated over a decade in the Narrative Layer and its built-in anti-synthesis mechanism—after the synthetic data route systematically fails in cross-validation between the Cognitive and Behavioral Layers, latecomer competitors must accumulate real data from scratch. The next phase of enterprise AI competition will shift from model access capabilities to domain understanding depth in business scenarios and the accumulation of technical assets.