🤖 AI

Virtual Personas for Language Models via an Anthology of Backstories

This paper introduces "Anthology," a method that conditions large language models on open-ended life narratives (backstories) to create virtual personas, demonstrating significant improvements in matching human response distributions and experimental consistency across three nationally representative surveys.

Original authors: Suhong Moon, Marwa Abdulhai, Minwoo Kang, Joseph Suh, Widyadewi Soedarmadji, Eran Kohen Behar, David M. Chan, John Canny

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Suhong Moon, Marwa Abdulhai, Minwoo Kang, Joseph Suh, Widyadewi Soedarmadji, Eran Kohen Behar, David M. Chan, John Canny

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a giant, super-smart robot that has read almost every book, article, and blog post ever written on the internet. Because it read so much from so many different people, this robot actually contains the "voices" of millions of different humans inside its brain.

However, there's a problem. When we usually talk to these robots (like ChatGPT), we've trained them to be helpful, polite, and neutral. It's like taking a room full of people from all walks of life—angry poets, shy accountants, excited teenagers, grumpy retirees—and forcing them all to wear the same boring gray suit and speak in a monotone voice. They lose their unique personalities.

This paper introduces a new method called "Anthology" to fix this. Here is how it works, broken down into simple steps:

1. The Problem: The "Gray Suit" Effect

Most researchers try to make the robot act like a specific person by giving it a list of facts, like: "You are a 35-year-old man named Bob who likes fishing."
The paper argues this is like giving an actor a script that just says "Be sad." The robot doesn't really feel sad; it just guesses what a sad person might say. It's stiff and often wrong.

2. The Solution: The "Life Story" (Backstory)

Instead of a list of facts, the researchers give the robot a backstory. This is a long, open-ended story written in the first person, like a diary entry or a memoir.

The Analogy: Imagine you want an actor to play a character. Instead of handing them a resume that says "Age: 40, Job: Teacher," you hand them a 10-page letter they wrote about their childhood, their divorce, their love for jazz, and their fear of heights.
The Result: When the robot reads this story, it doesn't just "know" the facts; it becomes the character. The story naturally includes their age, background, and personality without the robot having to guess.

3. How They Made the Stories (The "Anthology")

The researchers didn't ask real people to write these stories (which would take forever). Instead, they asked the robot itself to write thousands of these life stories.

They asked the robot: "Tell me about yourself."
The robot generated thousands of unique, detailed life stories, creating a massive library (an Anthology) of virtual people. Some were rich, some poor, some young, some old, some from the city, some from the country.

4. The "Matching" Game

Now they had a library of virtual people, but they needed to pick the right ones to match a real group of people (like a national survey).

The Analogy: Imagine you are casting a play and you need an audience that looks exactly like the population of New York City. You have a huge room full of actors (the Anthology). You don't just pick them randomly. You check their "demographics" (age, race, income) and use a smart matching system to pair each real person in your target group with the virtual actor who fits them best.
The paper uses two methods to do this matching: a "Greedy" method (pick the best match for each person immediately) and a "Maximum Weight" method (find the perfect overall balance for the whole group).

5. The Test: Did It Work?

The researchers tested this by having these virtual people take real surveys that real humans had taken (about topics like food, politics, and AI).

The Result: The virtual people created with the "Backstory" method answered the surveys much more like real humans than the old methods did.
The Numbers: They improved the accuracy of matching human responses by up to 18% and made the answers more consistent by 27%.
The Bonus: It worked especially well for groups of people that are usually hard to represent or study (like specific minority groups), because the robot could generate stories for them just as easily as for anyone else.

Why This Matters (According to the Paper)

The authors say this isn't just about replacing real people with robots. Instead, it's like giving scientists a new, powerful tool.

Safety: You can test ideas on these virtual people without risking harm to real humans.
Speed & Cost: You can ask 10,000 virtual people a question in seconds for free, whereas interviewing 10,000 real people costs a fortune and takes months.
Diversity: You can create a virtual crowd that perfectly represents a specific demographic, ensuring that no group is left out of the conversation.

Important Note from the Paper:
The authors warn that this only works if you use the "raw" version of the robot (the one before it was taught to be polite and helpful). If you use the "polite" version (the one we usually chat with), it refuses to act like a grumpy or complex human, and the method fails. They also warn that while this is a great tool for research, we must be careful not to use these fake people to trick real humans or spread misinformation.

In short: Anthology is a way to give AI a rich, detailed life story so it can stop sounding like a robot and start sounding like a real, diverse human being.

Technical Summary: Virtual Personas for Language Models via an Anthology of Backstories

Problem Statement
Large Language Models (LLMs) are trained on vast corpora containing the diverse voices of hundreds of millions of authors. However, standard applications of LLMs, particularly those involving instruction tuning (e.g., RLHF), suppress this inherent diversity to produce a single, helpful, and largely unemotional agent voice. This "alignment" makes it difficult to use LLMs as proxies for human subjects in behavioral studies, where representing a wide spectrum of human traits, demographics, and opinions is critical. Existing methods for conditioning LLMs to specific personas—such as providing free-text biographies or sequences of demographic question-answer pairs—have shown limited success in achieving high representativeness, consistency, and diversity, particularly for under-represented sub-populations. Furthermore, instruction-tuned models often fail to generate realistic, nuanced backstories, instead producing short, evasive, or overly positive responses that lack the complexity of human narratives.

Methodology: The Anthology Approach
The authors propose "Anthology," a methodology for conditioning LLMs to diverse, consistent, and representative virtual personas using open-ended life narratives, termed "backstories." The approach consists of four primary stages:

LLM-Generated Backstories: Instead of relying on limited existing autobiographical datasets, the authors use pre-trained (non-instruction-tuned) LLMs to generate a massive "anthology" of backstories. By prompting models with open-ended questions like "Tell me about yourself," they generate naturalistic, first-person narratives that implicitly encode demographic details (age, location, financial status) and explicit traits (values, personality, beliefs). The authors emphasize that pre-trained models are essential here, as instruction-tuned models tend to produce generic, safe, and inconsistent narratives.
Demographic Survey on Virtual Personas: Since backstories do not explicitly state all demographic variables, the authors perform a "demographic survey" on each generated persona. They prompt the backstory-conditioned model with standard demographic questions (e.g., age, education, income) and sample multiple responses to estimate a probability distribution of traits for that specific virtual subject.
Matching Target Human Populations: To approximate a specific human study, the authors match the virtual personas to the target human population's demographic distribution. They model this as a bipartite matching problem between human users ( $H$ $H$ ) and virtual users ( $V$ $V$ ). Edge weights are calculated based on the likelihood of a virtual user's demographic distribution matching a human user's traits. Two strategies are explored:
- Greedy Matching: Assigns each human user to the virtual persona with the highest similarity weight, allowing multiple humans to map to the same virtual persona.
- Maximum Weight Matching: Finds an optimal one-to-one correspondence to maximize the total weight, using the Hungarian algorithm.
Approximating Human Studies: The matched subset of backstories is used as context (prefix) for the LLM when answering survey questions. The study replicates the mental process of human respondents by presenting questions in series, including previous answers as context.

Key Contributions

Anthology Framework: Introduction of a method to generate and utilize an "anthology" of LLM-generated backstories to condition models, demonstrating that this approach yields more representative and consistent virtual personas than baseline methods (Bio and QA).
Demographic Matching: A formalized method for sampling and matching virtual personas to target human population distributions, significantly improving the approximation of human response distributions.
Open-Source Resources: Release of an open-source anthology of approximately 10,000 backstories and the associated code for generating, processing, and administering surveys.
Model Selection Insight: Demonstration that pre-trained models outperform instruction-tuned models for this task, as the latter have converged to a singular, homogenized persona that lacks the diversity required for behavioral simulation.

Experimental Results
The methodology was validated against three nationally representative surveys from the Pew Research Center's American Trends Panel (ATP Waves 34, 92, and 99), covering topics from biomedical issues to political typology and AI.

Representativeness: Anthology achieved up to an 18% improvement in matching human response distributions (measured by Wasserstein Distance) compared to baseline methods.
Consistency: The method showed up to a 27% improvement in consistency metrics (measured by the Frobenius norm of correlation matrices and Cronbach's alpha), indicating that the virtual personas maintained coherent internal logic across different survey questions.
Diversity: Anthology outperformed baselines across diverse demographic subgroups (race, age, education, gender). Notably, it achieved the lowest Wasserstein distances for under-represented groups, though the authors note that matching accuracy decreases for older age groups (65+), potentially due to the models' inherent bias toward younger demographics.
Matching Strategy: Greedy matching generally outperformed maximum weight matching in terms of Wasserstein distance, likely because the latter's one-to-one constraint forced suboptimal pairings given the finite pool of virtual users.

Significance and Claims
The paper posits that Anthology provides a new instrument for studying human behavior, distinct from a simple replacement for human studies. By leveraging the "upstream" diversity of pre-trained models and open-ended narratives, the method allows researchers to:

Simulate large-scale human studies at a fraction of the cost and time.
Access hard-to-reach or under-represented populations (e.g., homeless, incarcerated, or non-cooperative subjects) in a scalable manner.
Conduct "scrutable" studies where subjects can be queried in natural language about their reasoning.

The authors maintain a modest stance, acknowledging that while Anthology improves the fidelity of virtual subjects, it does not yet fully replicate complex human behaviors such as long-term attitude change or in-group/out-group biases. They highlight ethical concerns regarding the potential misuse of high-fidelity personas for social engineering or misinformation, and emphasize that the lower density of under-represented groups in training data remains a challenge for perfect simulation. Ultimately, the work suggests that carefully designed virtual personas can serve as valuable adjuncts to human research, offering a scalable tool for exploring the "enormous diversity of human traits" embedded within language models.