Xinmin Evening News | Subjective World Model Allows AI to 'Understand Human Hearts'

Professor Fan Ling of Tongji University, Director of the Design Artificial Intelligence Laboratory, Founder of Tezign Technology

This is an original AI framework from China—the Subjective World Model. It differs from the overseas LLM route, which focuses on text prediction, and carves out a path of AI that 'understands people'. It upgrades human-computer interaction: moving from 'Q&A' to 'empathetic understanding', allowing AI to begin to know what people are struggling with, where their words and feelings diverge, and ultimately how they will choose. However, this is still simulation, not AI truly possessing subjective consciousness.

Can AI truly understand people? It is not about 'understanding your instructions', nor 'responding based on your input'—but understanding you as a person: 'Why do you think this way? What do you truly care about?' 'The gap between what you say and your real priorities.'

A 28-year-old white-collar woman in Shanghai wrote on social media: 'I am a rational consumer, I only look at ingredients, not brands.' But when her circle of friends was using a certain luxury brand, she bought it. Afterwards, she posted a note recommending 'ingredient-conscious products', explaining that this was 'rational judgment'. What people say, explain, truly care about, and ultimately do—these four layers are often filled with contradictions. But this is precisely what makes a real person.

Technically Feasible to 'Understand People'

In the past three years, the capabilities of large language models have continuously expanded—from generating text, images, and videos, to reasoning, writing code, and autonomously completing complex tasks. However, all these models share the same training goal: predicting the probability distribution of the next word. This makes them extremely adept at understanding and generating language, yet they have always struggled with one thing—understanding the speaker themselves.

A simple counterexample: if you ask the same large language model 'Do you prefer coffee or tea?', the answer today and tomorrow may be completely different. Because the model has no 'you', only the current context. It understands language but does not model people.

Philosopher Wittgenstein famously said, 'The limits of my language mean the limits of my world.' Conversely, if one could fully describe a person using language, it would be equivalent to entering their world. So, is there a model that does not aim at language distribution but specifically models the subjective world of people—those thoughts, preferences, struggles, and contradictions?

The academic community has already provided a preliminary answer. A research team from Stanford University and Google constructed 1,000 'digital avatars' based on large language models, allowing AI to answer questionnaires on behalf of real people. The results were surprising: the consistency rate between AI avatars and real human responses reached 85%—while the same person redoing the same questionnaire two weeks apart had a consistency rate no higher than this. Another study found that AI has quantifiable 'personality' dimensions that can be detected, predicted, and even precisely adjusted. These studies collectively point to a possibility: modeling the subjective world of people based on language models is technically feasible.

Four Layers of Data Collaborative Training

At the recently concluded 2026 World Artificial Intelligence Conference, a new model framework attracted significant attention—the Subjective World Model (SWM). This model was first developed by Tezign Technology.

The fundamental difference between SWM and existing large language models lies in the different training objectives: large language models optimize the fitting of language distributions; SWM optimizes the accuracy of behavior prediction. To this end, it constructs a four-layer progressive heterogeneous data training architecture, with each layer corresponding to the real state of a person at different depths:

■ Expression Layer collects billions of pieces of native social media data, modeling the mapping relationship between language style and user characteristics. This layer captures what people say in public—style, expression of preferences, and construction of personas.

■ Story Layer collects tens of thousands of hours of one-on-one in-depth interview data, constructing a temporal causal chain of behavioral motivations. This layer captures how people explain their choices and the complete story from triggering to decision-making and post-hoc rationalization behind that explanation.

■ Cognition Layer introduces psychological standard scales such as Schwartz Value Survey and Big Five Inventory, systematically correcting the deviation between 'stated preferences' and 'true preferences'. This layer touches on what people truly care about—often different from what they verbally express.

■ Behavior Layer estimates individual loss aversion coefficients, time discount rates, and sensitivity to social norms through economic experiments like ultimatum games and public goods games, predicting what people will ultimately do—actual decision-making in situations with real stakes.

The collaborative training of these four layers outputs an AI persona with a persistent internal state. It can maintain internal consistency when continuously questioned, unlike ordinary language models that generate contradictory answers each time—because it maintains not a context window, but a person's motivational structure and behavioral parameters.

'Stated Preferences' Cannot Be Fully Trusted

To understand the significance of SWM, one must first confront a fundamental issue: what people say and what they do are often two different things.

Behavioral economics refers to this phenomenon as the systematic bias between 'stated preferences' and 'revealed preferences'. There are numerous cases in consumer research: people claim to value environmental protection in questionnaires, but in actual purchasing behavior, the weight of price factors far exceeds that of environmental factors; people say they consume rationally, but under the influence of their social circles, purchasing decisions frequently deviate from rational assumptions.

What people say, explain, truly care about, and ultimately do—these four layers are filled with contradictions. But this is precisely what makes a real consumer.

Traditional questionnaires can only capture the expression layer; in-depth interviews can reach the narrative layer; but the value weight bias of the cognition layer and the true decision parameters of the behavior layer can only be measured through specially designed experiments and real behavioral data. This is the core of SWM's four-layer architecture that surpasses existing methods.

Deep Dialogue Layer by Layer Inquiry

The commercial application of SWM is realized through Tezign Technology's user research platform Atypica. It transforms AI Persona into a repeatedly callable consumer cognition system—validating any product, any idea, at any moment, instantly. This model is further applied in Tezign's enterprise-level intelligent product series GEA, serving as the core technology that enables large language models to truly understand target consumers.

A maternal and infant company once faced this confusion: young mothers frequently claimed on social media to 'only buy cost-effective products', but purchasing data showed they repurchased the most expensive products. Traditional questionnaires could only capture the content users were willing to express.

Atypica's approach is different. It first engages AI in deep dialogue with 2,000 real users—not structured questionnaires, but layer-by-layer inquiries like chatting with friends, until the true decision logic is touched. The results clearly show: mothers say 'cost-effectiveness', but when their mothers say 'you can't skimp on children's things', their choices immediately flip.

This is no longer a report that gets archived after reading; it is a living consumer cognition system that can be repeatedly called upon. In the past, conducting a survey of a thousand people required weeks of time and a six-figure budget; through Atypica, the speed increases by 100 times, costs decrease by 100 times, and the reachable consumer sample size expands by 100 times.

The Boundaries of a New Paradigm

The Subjective World Model did not emerge out of nowhere. It stands on the achievements accumulated in the past three years in the academic community in cognitive modeling, behavioral modeling, and personality modeling, proposing an engineering implementation path: using heterogeneous multi-layer data for collaborative training, replacing the single language corpus pre-training objective. This is a fundamental expansion of the large language model paradigm, not just an improvement at the application level.

The significance of this paradigm may far exceed the commercial applications themselves. If AI can reliably model the subjective cognitive structures of people, it becomes not just a production tool, but a scientific instrument for understanding human behavior—potentially opening up new possibilities in public policy simulation, social science research, and mental health interventions.

Of course, this direction also faces profound challenges: can training data truly reflect cognitive differences under diverse cultures? How can the model's generalization ability beyond the sample be validated? Where is the boundary between commercial applications and personal privacy protection? These questions still do not have complete answers. Precisely because of this, they deserve to be raised seriously.

AI understanding people is not to manipulate them more precisely. If the endpoint of this technology is to bring business closer to people, to make products respond more authentically to human needs, rather than amplifying biases and exacerbating manipulation—then it requires not just engineers, but also psychologists, sociologists, ethicists, public participation, and regulatory intervention.

The Subjective World Model is a beginning. It proposes a direction: AI is not just a production tool, but can also be a tool for understanding people. But 'understanding' itself comes with responsibility. We hope to participate in establishing the boundaries it should have while advancing this technology.

(Source: Xinmin Evening News, Reporter Zhang Jiongqiang)