💬 NLP

Anamnesis: An Open-Source Platform for Large-Scale Backstory-Conditioned Survey Simulation

This paper introduces Anamnesis, an open-source, interactive platform that leverages large language models conditioned on structured narrative backstories to simulate large-scale, demographically controllable surveys, demonstrating its ability to more accurately replicate real-world opinion distributions than standard persona-prompting baselines.

Original authors: Song-Ze Yu, Joseph Suh, Serina Chang, David M. Chan

Published 2026-07-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Song-Ze Yu, Joseph Suh, Serina Chang, David M. Chan

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 want to know what a whole city thinks about a new park, but you can't ask every single person. Traditionally, researchers have to pay real people to fill out surveys, which is slow, expensive, and sometimes hard to organize. Enter Anamnesis, a new open-source playground that lets researchers build "virtual populations" to test their questions before they ever talk to a real human.

Think of Anamnesis as a massive, digital casting call. Instead of hiring actors who just say, "I am a 35-year-old woman," the system casts characters with full, rich life stories. These aren't just lists of facts; they are detailed, multi-paragraph backstories filled with memories, values, and quirks, generated by AI. When you ask these virtual characters a survey question, they answer based on their entire "life" so far, not just a label.

Why the old way didn't work as well
The paper argues that the old method of "persona prompting"—where you just tell an AI, "Act like a 35-year-old Hispanic woman"—is like asking a method actor to play a role with only a name tag. It often leads to stereotypical answers that feel flat and fake. The authors show that these simple prompts fail to capture the messy, nuanced, and sometimes contradictory opinions of real people. Anamnesis rejects this shallow approach, insisting that to get real opinions, you need deep, narrative backstories.

How the magic happens
The platform works like a sophisticated matchmaking service.

The Casting Pool: The system has a database of 35,000 pre-made virtual characters, each with a unique life story.
The Target: A researcher might say, "I need 100 people who are voters aged 25–44, split evenly between Democrats and Republicans."
The Match: Because the system doesn't know the exact demographics of every character with 100% certainty (it's an estimate), it uses a clever math trick to pick the best group that fits the researcher's target. It's like a DJ mixing tracks to get the perfect vibe for a party, ensuring the crowd feels just right.
The Conversation: As the virtual people answer the survey, the system remembers their previous answers. If a character says they love hiking in question one, they won't suddenly say they hate the outdoors in question two. This keeps their "personality" consistent, just like a real human would be.

Did it actually work?
The authors didn't just guess; they put the system to the test in two specific ways:

The Political Test: They tried to recreate real surveys from the Pew Research Center about topics like politics and biomedical issues. The results showed that the Anamnesis virtual population gave answers that were much closer to real human data than the old "short list" prompting methods. In fact, the virtual group's opinions were so similar to real people that the difference was measured as a 0.147 (a low number means a close match) in one specific test, compared to 0.258 for the older, simpler method.
The Cartoon Test: They also tested if these virtual people could understand humor and images. They used the New Yorker Caption Contest, where people vote on the funniest caption for a cartoon. When the virtual characters voted, they picked the human-favorite captions 59.2% of the time, compared to only 51.0% for the simpler method. This suggests that giving the AI a full life story helps it "get" human preferences better, even with pictures and jokes.

What this means for you
This isn't a magic crystal ball that predicts the future, nor is it a replacement for real human research. The authors suggest that Anamnesis is a powerful prototype tool. It allows researchers to "stress-test" their surveys—checking if their questions are confusing or if their results look weird—using a diverse crowd of virtual people first. If the virtual crowd reacts strangely, the researcher can fix the survey before spending money on real people.

The paper emphasizes that this is an open-source tool, meaning anyone can use it for free, unlike some commercial "black box" services that hide how they work. By making these advanced simulation techniques available to everyone, Anamnesis hopes to help researchers understand human behavior more accurately, faster, and cheaper, all while keeping the process transparent and reproducible.

Technical Summary: Anamnesis

Problem Statement

Traditional opinion surveys and social polling face significant logistical and economic hurdles, including rising costs, declining response rates, and the difficulty of reaching specific demographic sub-populations. While Large Language Models (LLMs) offer a potential alternative as "virtual personas," previous attempts to simulate human populations have relied on "persona prompting" using short lists of demographic attributes (e.g., "35-year-old Hispanic woman"). This approach often yields stereotypical responses and lacks the psychological depth required for complex opinion elicitation. Furthermore, existing high-fidelity methodologies like Anthology (Moon et al., 2024) and Alterity (Kang et al., 2025) remain confined to siloed Python scripts, while commercial simulation platforms operate as closed-source, proprietary systems lacking the transparency and reproducibility necessary for rigorous social science.

Methodology

Anamnesis is an open-source, web-based platform designed to operationalize the Anthology and Alterity frameworks for non-technical users. It replaces short demographic prompts with rich, open-ended narrative backstories to condition LLM responses.

Core Components

Narrative Backstory Conditioning:
- Instead of attribute lists, the system conditions models on multi-paragraph life histories generated by sampling from pretrained base models (using interview questions from the American Voices Project).
- These backstories capture formative experiences, values, and worldviews, not just demographics.
- Demographic attributes are inferred probabilistically from the backstories rather than being observed labels.
System Architecture:
- Dispatcher–Queue–Worker Model: To handle the $O(S \times Q \times N)$ complexity of survey runs (where $S$ is sample size, $Q$ is questions, and $N$ is resampling), the system uses an asynchronous queue. Workers execute questions sequentially per persona with sequential context accumulation, ensuring cross-question belief consistency.
- Survey Builder: A graphical interface allows users to construct multi-modal surveys (multiple-choice, open-ended, ranking, image, and audio) without coding.
Demographic Targeting & Selection:
- Probabilistic Matching: Since demographics are inferred distributions ( $p_{b,d}(c)$ $p_{b, d} (c)$ ), the system employs two selection algorithms:
  - Top-K Probability Ranking: Selects personas with the highest joint compatibility with target filters.
  - Balanced Demographic Matching: Solves an assignment problem (using the Hungarian algorithm) to fill specific demographic cells (e.g., age $\times$ gender) while maximizing overall compatibility. To ensure real-time latency, the candidate pool is pruned to the top- $M$ candidates per cell before assignment.
- Demographic Inference: Users can define new demographic dimensions (e.g., political leaning) on the fly. The system estimates probability distributions for these dimensions using either Logprobs mode (single constrained forward pass) or N-sample mode (empirical distribution from repeated sampling) to accommodate users without self-hosted vLLM infrastructure.
Answering Algorithms:
- Anthology (Default): Prepend the full backstory to the first question; append question-answer pairs to the context for subsequent questions.
- Zero-shot Baseline: Conditions only on short demographic descriptions (e.g., CLAIR-style prompts) to serve as an ablation control.

Key Contributions

First Open-Source GUI Platform: Anamnesis is the first system to combine narrative backstory conditioning, probabilistic demographic matching, sequential context accumulation, and multimodal survey support in a deployable, non-technical interface.
Operationalization of Advanced Frameworks: It translates the Anthology and Alterity methodologies from research prototypes into a unified web interface, making state-of-the-art persona approximation accessible to researchers without programming expertise.
Scalable Infrastructure: The platform supports concurrent survey runs over a pool of 35,000 pre-sampled backstories with automated demographic balancing and result aggregation.

Results

The system was evaluated through two case studies:

Replicating Public Opinion Polls (Pew Research Center ATP):
- The platform replicated segments of the American Trends Panel (ATP) across three waves (biomedical, political typology, AI).
- Metric: Average Wasserstein Distance (WD) and Frobenius Norm (Fro.) between response correlation matrices.
- Outcome: Anamnesis (backstory-conditioned) significantly outperformed standard demographic-list baselines (BIO and QA) in matching real-world response distributions. For example, in Wave 34, Anamnesis achieved a WD of 0.147 compared to 0.258 for the BIO baseline. The results closely mirrored human distributions, though a gap remained compared to actual human data.
Multimodal Preference Estimation (New Yorker Caption Contest):
- The system simulated human vision-language preferences on 49 cartoon-caption contests.
- Metric: Majority-vote accuracy and vote share for the human-preferred caption.
- Outcome: The Anthology approach achieved 59.2% majority-vote accuracy (vs. 51.0% for the zero-shot baseline). It increased the mean vote share for the human-preferred caption by 7.8 percentage points ( $p=0.0024$ ), demonstrating that narrative conditioning improves alignment with collective human judgments in multimodal tasks.

Significance

The paper positions Anamnesis as a transparent, reproducible, and open-source alternative to proprietary simulation services. Its primary utility lies in enabling social scientists to prototype and stress-test survey instruments on diverse, demographically balanced virtual populations before recruiting human subjects. By providing a non-technical interface for high-fidelity persona simulation, the platform aims to democratize access to rigorous social science methodologies, allowing researchers to examine opinion distributions across subgroups without the logistical burden of traditional human-subject research.