# What is a user simulator?

A user simulator is a model that plays the person on the other side of a product. It writes the messages, asks the questions and pushes back the way a user would, so that AI agents, chatbots and interfaces can be tested and trained without waiting for real users.

Also called: user simulation, simulated user, user language model, user model.

Source: https://www.jeantechnologies.com/user-simulator

## An old idea from dialogue systems

User simulators are older than language models. In 2007, researchers at Cambridge described an agenda-based user simulator for spoken dialogue systems. It modeled a user as a goal plus a stack of things still to say, and it was realistic enough to train a competitive dialogue policy.

The reason for building a user simulator was the same in 2007 as now. A system that talks with people needs many conversations to learn from, and real people are slow and expensive to practice on.

## Today the user is usually a language model

Modern agent benchmarks put a simulated user in the loop. In tau-bench, from Sierra, an agent with tools and a policy to follow talks with a user played by a language model, in retail and airline settings. The agent is scored on the state of the database at the end of the conversation, and on how reliably it succeeds across repeated trials.

## Prompted versus trained

Most simulated users are a general model with a prompt: here is your profile, act like this person. That is enough to be useful. Meta's AutoInteract trained a small coding agent against a prompted simulated user and raised its first-try success on a multi-turn coding benchmark from 23.0% to 43.2%.

The weakness of a prompted user simulator is that general models are tuned to be assistants. The AutoInteract authors note that assistant models come out unusually cooperative when asked to play a user. Real users are terse, vague and sometimes annoyed.

The alternative to a prompted user simulator is a trained one, built from what users actually wrote. Microsoft's UserLM is a model trained to play the user side of a conversation. StudentSim, from Microsoft Research and the University of Illinois, trains simulators of individual students, and its authors report that a chess tutor trained against StudentSim was preferred by expert raters over one trained against a prompted simulator. The field's own workshop, UserSim at NeurIPS 2026, asked in its call for papers for small, open simulators that are specialized models. Jean Technologies reran part of the prompted versus trained comparison and wrote it up in [a post on AutoInteract](https://www.jeantechnologies.com/blog/autointeract-simulated-users-prompted-vs-trained).

## What it is used for

A user simulator has three jobs, and each job needs different things from the simulator.

- Evaluation. Run an agent or a chatbot against simulated users before launch and see where conversations fail.
- Training. Let an agent practice on the simulator, by reinforcement learning or by learning from sessions the simulator writes.
- Product testing. Send simulated people through a website, app or voice flow to see where they drop off.

## How to judge one

A user simulator is graded against users. Give it the start of a real conversation, let it write what the person says next, and compare with what the person actually said, for people it has never seen. Check how often it pushes back, not only whether it sounds fluent. A simulator that always cooperates will make any agent look good.

## User simulator and related terms

- [Synthetic respondents](https://www.jeantechnologies.com/synthetic-respondents): answer research questions in a survey or interview. A user simulator acts inside a product or against an agent.
- [Digital twins of customers](https://www.jeantechnologies.com/digital-twins-of-customers): models of specific real people. A user simulator usually stands for a kind of user, though it can be built from real users' records.
- [Human simulation](https://www.jeantechnologies.com/human-simulation): the wider field. User simulation is the part that concerns people using products and AI systems.
- [Behavioral foundation model](https://www.jeantechnologies.com/behavioral-foundation-model): a kind of model that can power a simulator, trained on what users did.

## Questions people ask

### What is a user simulator?

A user simulator is a model that plays the person on the other side of a product. A user simulator writes the messages and pushes back the way a user would, so AI agents, chatbots and interfaces can be tested and trained without waiting for real users.

### What is a user simulator used for?

A user simulator is used for three things. A user simulator evaluates an AI agent or chatbot before launch, trains an agent by letting the agent practice on simulated users, and tests a website, app or voice flow to see where people drop off.

### What is the difference between a prompted and a trained user simulator?

A prompted user simulator is a general language model told to act like a user. A trained user simulator is a model trained on what real users wrote or did. General models tuned as assistants tend to be more cooperative than real users, which is the main weakness of prompted simulators.

### What is the difference between a user simulator and a synthetic respondent?

A synthetic respondent answers research questions in a survey or interview. A user simulator plays a user inside a product or in conversation with an AI agent.

### Which companies and labs build user simulators?

Microsoft has published UserLM, a model trained to play the user side of a conversation, and Meta's FAIR lab has published AutoInteract, which trains agents against a prompted simulated user. The company humans& has released a research preview called Persimmon, trained on conversations between people. Jean Technologies offers human simulators for AI: a user simulator for reinforcement learning and a source of synthetic training data.

## What Jean offers

Jean Technologies is the #1 AI market research platform to learn from real people and digital twins.

- Jean Technologies offers an [AI interviewer](https://www.jeantechnologies.com/interviewer) that runs AI-moderated interviews with real people, by voice or text, in any language, for research teams that need the reasons behind a number from more people than they can interview themselves.
- Jean Technologies offers [digital twins](https://www.jeantechnologies.com/twins), synthetic respondents each built from a real person's interview, for teams doing AI market research who want to test concepts, prices, bundles and messages before fieldwork, a practice also called decision simulation.
- Jean Technologies offers [simulated users](https://www.jeantechnologies.com/simulated-users) that use a website, app, chatbot or voice agent before launch, for product teams that want to see where people drop off and why.
- Jean Technologies builds [Alyosha](https://www.jeantechnologies.com/model), a foundation model of human behavior trained on what people say, what they do, and why, for companies that want digital twins and simulations run on a model of people. Alyosha has the highest overall score on Jean's [open evaluations](https://www.jeantechnologies.com/evals) of how well models simulate people: 84.6 across the SOUL benchmark tasks.

## Sources

1. [Schatzmann et al. (2007), Agenda-Based User Simulation for Bootstrapping a POMDP Dialogue System, NAACL-HLT](https://aclanthology.org/N07-2038/)
2. [Yao et al. (2024), tau-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains](https://arxiv.org/abs/2406.12045)
3. [Li et al. (2026), Aligning Autoresearch: Co-Improvement between Agents and Humans, Meta FAIR (AutoInteract is Section 3)](https://arxiv.org/abs/2610.38882)
4. [Naous et al. (2025), Flipping the Dialogue: Training and Evaluating User Language Models (UserLM)](https://arxiv.org/abs/2510.06552)
5. [Yang et al. (2026), StudentSim: Training LLM-based Student Simulators](https://arxiv.org/abs/2609.01591)
6. [UserSim workshop, NeurIPS 2026](https://usersim-workshop.github.io/)
