Introduction
Large language models (LLMs) are trained on immense text corpora—the collective output of millions of diverse human authors. This raises an intriguing question: can we guide an LLM to embody a specific individual's voice rather than the default blend of voices? Recent research, particularly in "Language Models as Agent Models," shows that LLMs can act as models of agents: given the right context, they produce text that reflects a particular agent's characteristics. This capability opens the door to cost-effective virtual personas for social sciences and user research, aligning with ethical principles like justice and beneficence. However, early attempts using simple demographic prompts (e.g., "I am a 25-year-old from California") often yielded stereotypical responses and failed to capture individual-level variation. Enter Anthology: a method that conditions LLMs with richly detailed life narratives—backstories—to produce representative, consistent, and diverse virtual personas. This guide walks you through implementing Anthology to generate and use backstories for creating authentic virtual subjects.
What You Need
- An LLM with strong generative capabilities (e.g., GPT-4, Claude, or open-source models like Llama 2) – The model must be capable of producing coherent, lengthy narratives.
- A demographic taxonomy – A set of variables describing human attributes: age, gender, education level, occupation, geographical region, cultural background, etc.
- A backstory generation script – Code to prompt the LLM systematically, feeding demographic tuples to produce individual life narratives.
- A conditioning framework – The ability to prepend generated backstories as context for subsequent LLM queries (e.g., via prompt engineering or API parameters).
- Evaluation metrics – Tools to assess representativeness (match with human response distributions), consistency (stability across multiple queries), and diversity (coverage of demographic groups).
Step-by-Step Instructions
Step 1: Define Your Demographic Space
Start by specifying the demographic variables you want your virtual personas to represent. For Anthology to work effectively, these should be more nuanced than a simple tuple—consider life experiences, values, and socioeconomic factors. For example:
- Age (e.g., 25, 45, 70)
- Gender identity
- Highest education level
- Occupation or industry
- Geographic location (city/state/country)
- Household income range
- Cultural or religious background
Create a list of such tuples—each tuple will seed one virtual persona.
Step 2: Generate Naturalistic Backstories
Use the LLM itself to generate a unique backstory for each demographic tuple. Write a prompt that instructs the model to create a first-person life narrative incorporating all the given variables. For instance:
"You are a 45-year-old woman from rural Texas with a high school education working as a waitress. Your family values are traditional. Write a 300-word story about your life experiences, including key events that shaped your beliefs and daily routine."
Ensure the output is rich—include details about values, challenges, and aspirations. This backstory becomes the conditioning context. Repeat for all tuples.
Step 3: Validate Backstory Quality
Not all generated backstories will be equally useful. Filter out those that feel generic or fail to incorporate the given demographics. A good backstory should:
- Be logically consistent (e.g., a history teacher from 2024 shouldn't mention owning a smartphone in 1985).
- Reflect nuanced, non-stereotypical details (avoid pure caricatures).
- Be long enough to provide meaningful context (≥150 words).
Optionally, have a human reviewer rate a sample of backstories for authenticity.
Step 4: Condition the LLM with Backstories
For each virtual persona you want to simulate, prepend its backstory to the LLM prompt. For example:
Context: [Full backstory text]
Now answer the following question from the perspective of the person described above:
Question: [Survey item or user research question]This tells the LLM to respond as that specific individual. Crucially, the backstory provides implicit information beyond demographics—values, lived experiences, cognitive biases—that the model can leverage to produce more human-like responses.
Step 5: Generate Virtual Responses
Run the conditioned LLM on your target questions (e.g., opinion surveys, behavioral tasks, open-ended interviews). For statistical power, generate multiple responses per persona—this allows you to compute within-person variance and aggregate at the individual level. Record all outputs.
Step 6: Evaluate Fidelity
Compare your virtual responses against real human data (if available) or against expected distributions. Key metrics include:
- Representativeness: Does the distribution of virtual responses match the distribution of human responses for the same demographic group? Use measures like Jensen-Shannon divergence or Kullback-Leibler divergence.
- Consistency: Does the same persona give similar answers to the same question repeated? Compute intraclass correlation coefficient (ICC) or Cronbach's alpha.
- Diversity: Are the backstories and resulting responses varied across demographic groups? Avoid mode collapse where all personas sound alike.
Adjust your generation prompts or filtering thresholds if fidelity is low.
Step 7: Scale and Iterate
Anthology's real power comes from scale—generate backstories for hundreds or thousands of personas covering a wide demographic space. Automate the generation and evaluation pipeline using scripts. Iterate on your demographic taxonomy and backstory prompts based on fidelity results.
Tips and Best Practices
- Diversity over quantity: Focus on covering a broad range of demographic combinations rather than generating many near-identical profiles. Aim for at least 20 distinct personas per key demographic variable.
- Include contradictory values: If your research involves sensitive topics, explicitly include value conflicts in backstories (e.g., a liberal Catholic) to avoid the LLM defaulting to stereotypes.
- Use temperature sampling: When generating backstories, set the LLM's temperature between 0.7 and 0.9 to encourage creative variety without coherence loss.
- Check for bias: Run bias checks on generated backstories—if certain groups consistently receive more negative narratives, adjust your prompt phrasing.
- Refer back to Step 4 when tuning conditioning: sometimes a shorter backstory works better than a long one; experiment with truncation.
- Document everything: Keep a log of demographic tuples, backstory texts, and evaluation results. This supports reproducibility and transparency.
- Combine with pre-existing surveys: For maximum realism, pair virtual personas with established survey instruments validated on real populations.
- Pilot test ethics: While Anthology can reduce the cost of human subjects research, never use virtual personas as a complete substitute for human input in high-stakes studies. Always follow the Belmont principles of justice and beneficence.