AI Agents Keep Converging on the Same Opinion. This Research Tries to Stop It

The promise of AI-based public-opinion research is easy to understand. A language model can answer hundreds of survey questions in minutes, repeat an experiment at low cost and let researchers test scenarios that would be slow or expensive with human participants.

The danger is just as important: a thousand simulated people can still be one model wearing a thousand name tags.

A KDD 2026 paper from researchers at Tsinghua University and Quancheng Laboratory gives that problem a precise name: diversity collapse. When large language models simulate people from different backgrounds, the agents may begin with distinct descriptions but converge toward similar internal representations and answers as information moves through the network. The result can look like a population while behaving like a compressed average.

The researchers propose a method called Parametric Social Identity Injection, or PSII. Instead of relying only on a written persona—age, education, income, religion and other attributes—it injects learned demographic and value-orientation vectors directly into selected hidden layers of an open model. Carefully calibrated noise is added to preserve variation among people who share the same broad attributes.

Across four open-source language models and hundreds of World Values Survey questions, the reported improvement is substantial. But the paper's most important message is not that synthetic agents can replace surveys. It is that realistic social simulation requires more than a convincing prompt, and that better statistical fit must still be separated from genuine understanding of human beings.

Why prompt-based populations become too tidy

Most synthetic-population systems begin with a profile in text: imagine a 45-year-old teacher, or answer as a rural voter with a particular education and income. That approach is intuitive because it resembles instructions a human interviewer might use.

The paper argues that it is also fragile. A persona exists outside the model as text. As the model processes the prompt through successive transformer layers, that identity signal can be smoothed away by the same machinery that favors common, probable continuations. Less frequent views may disappear, while different agents drift toward a small set of safe answers.

The authors visualize this effect using final-token hidden states from 500 simulated agents. Representations spread in intermediate layers, then contract again near the top of the network. Their label, diversity collapse, describes both flattened differences between groups and insufficient variation within them.

Those two failures pull in opposite directions. If identity information is too weak, groups become indistinguishable. If a prompt makes a demographic label too dominant, every member of that group begins to sound alike and stereotypes become stronger. A credible simulation needs to preserve between-group patterns without pretending that a demographic category determines an individual.

Simply raising the sampling temperature does not solve the problem. Randomness can produce different answers, but difference alone is not evidence that the output resembles a real population. The target is structured heterogeneity: variation that is calibrated against observed human response distributions.

What PSII changes inside the model

PSII combines three kinds of identity information.

First, each agent still receives a conventional written profile. This gives the model semantic context in familiar language.

Second, the system constructs vectors for demographic attributes. In simplified terms, researchers compare hidden-state patterns associated with different attribute values and derive directions that can steer the model. Rather than mentioning identity once at the beginning of a prompt, the system can reinforce it inside the network at a layer chosen for that attribute.

Third, the method adds value-orientation vectors built from text in five languages—English, Chinese, Arabic, Spanish and Russian. These are meant to capture broader cultural and normative signals beyond a demographic checklist.

The injection is hierarchical. Attributes associated with practical constraints and life circumstances can be placed in lower or middle layers; value-oriented information can be introduced higher in the network, where more abstract judgments are formed. The researchers first tested attributes across layers of Qwen2.5-7B and selected the layer that minimized divergence from human response distributions.

Finally, PSII applies model-specific Gaussian noise to demographic vectors. This is not random decoration. The noise level is calibrated to the model's sensitivity so that agents sharing a profile do not become identical while the intervention remains stable.

That architecture matters because it changes the objective from “write a more detailed character description” to “maintain controlled differences throughout generation.” It also limits where PSII can be used. Researchers need access to a model's internal activations and must calibrate interventions for each architecture; this is not a prompt that can simply be pasted into any closed chatbot.

The reported results are large

The evaluation uses the World Values Survey, a long-running cross-national research program covering beliefs, institutions, politics, economic attitudes and demographic information. The paper treats questions Q1 through Q259 as opinion targets and uses the remaining demographic questions to construct profiles and identity signals.

From a dataset of 97,220 respondents, the researchers sample groups of 100 people to create simulated populations. Each agent answers one fixed-choice question at a time. The experiment covers Qwen2.5-7B-Instruct, Qwen2.5-14B-Instruct, Llama-3.1-8B-Instruct and Mistral-24B-Instruct.

Performance is measured in two ways. KL divergence tracks how far the simulated answer distribution is from the human distribution; lower is better. Entropy Deviation measures whether the diversity of answers resembles the human benchmark; again, lower is better.

On Qwen2.5-7B, direct prompting produced an overall KL divergence of 1.3915 and Entropy Deviation of 0.7340. With PSII, those figures fell to 0.4843 and 0.0319. On Llama-3.1-8B, overall KL fell from 1.2856 to 0.4017, while Entropy Deviation dropped from 0.6326 to 0.0040.

PSII also outperformed the reported overall baselines on the other two models, although the size of the advantage varied. The 14-billion-parameter Qwen model showed a smaller diversity gain than its 7-billion-parameter sibling, underscoring that hidden-state steering is sensitive to architecture, depth and noise calibration.

An ablation study offers a useful reality check. Removing any major component reduced distributional accuracy. Demographic vectors were the most important for the combined accuracy-and-diversity result, while injected noise was especially important for preventing homogenization. Written profiles improved alignment but could narrow the output space, illustrating the tension between making an agent specific and making a population diverse.

The authors also resampled five groups of 100 respondents. PSII's overall metrics remained relatively stable in that limited robustness test. That is encouraging, but it is not the same as independent replication across survey waves, countries, model families or real-world events.

What the experiment does not prove

A model that matches an answer distribution has not demonstrated that it understands why people hold those views. Statistical resemblance can come from useful representation learning, from accidental correlations or from encoding the biases already present in the survey and model.

The task is also unusually constrained. Agents select among fixed survey options. Real public opinion is expressed through stories, arguments, uncertainty, social pressure and change over time. A method that improves a closed-choice benchmark may not preserve diversity in an open debate or accurately predict reactions to a new event.

The simulated population is built from coarse attributes. The authors acknowledge that rare or underrepresented groups remain limited by the data available, while personality, changing opinions and context-specific behavior are not fully modeled. Turning demographic differences into steering vectors creates a further risk: the system may make group averages feel like individual destiny.

The paper explicitly restricts PSII to aggregate, controlled research. It should not be used to profile individuals or make real-world decisions about them. The authors also warn that large synthetic populations could be misused for propaganda, social manipulation or the mass production of plausible narratives.

This distinction matters for businesses and policymakers. A simulation can be a hypothesis generator—a way to identify questions, stress-test message assumptions or explore potential distributions. It is not public consent, a replacement for fieldwork or proof of what a particular community believes.

Why this matters beyond opinion polling

The diversity-collapse problem appears wherever many agents are asked to stand in for a varied human group. Product research, policy modeling, classroom simulations and multi-agent debate can all produce the illusion of breadth while recycling the preferences of one underlying model.

PSII suggests that better agent design may require interventions below the prompt layer. Persistent internal controls can outperform elaborate character biographies because they operate where the model forms its answer rather than only where it receives instructions.

The broader lesson is methodological. Synthetic populations should be judged as distributions, not as collections of persuasive anecdotes. Researchers need human benchmarks, explicit diversity measures, ablations and tests for sensitivity to sampling. A few believable transcripts are not validation.

There is also a governance lesson. The more controllable simulated identities become, the easier it is to study representation—and to manufacture it. Access controls, audit logs, disclosure of synthetic evidence and strict prohibitions on individual profiling should develop alongside the technical capability.

What to watch next

The official code and identity vectors have been released, making independent inspection possible. The next test is whether other teams can reproduce the results and extend them beyond one survey framework without amplifying stereotypes.

Three questions deserve attention. Does the method still work on open-ended responses? Can it model opinion change after new information rather than a static answer? And can diversity improve without embedding demographic categories so strongly that they become caricatures?

PSII is a meaningful attempt to make simulated publics less artificial. Its success should make researchers more demanding, not less: a synthetic crowd must preserve disagreement, uncertainty and variation before anyone treats it as evidence about society.

Signal & Syntax will continue to update this guide as products, access and practical evidence change.