Noema

Noema is an experimental creative AI. It uses dynamic persona generation as a mechanism for inducing diverse creative trajectories in a fixed LLM.

Instead of relying on a single personality, Noema randomly generates a persona from a seed and applies that to the user's request.

The goal is to explore whether changing the perspective of a model can produce varied creative results, hopefully to be a creative tool.

Installation

Noema isn't intended to be a product. However, there are Releases made right here if you wish to experiment with the tool.

Required prerequisites

• Windows OS (or an emulation)
• .NET 10 Runtime
• Ollama with at least one installed model

How to use it

• Unzip the latest build into somewhere safe.
• Copy the path of the build
• Open cmd/Powershell and type cd PATH_LOCATION (pasting is right click)
• Then run with .\Noema.exe --model YOUR_CHOSEN_MODEL

If you don't want to cd to the .exe every time, there's a solution.

• Unzip the latest build into somewhere safe.
• Copy the path of the build
• Open the search bar and type Environment Variables
• Click on the Environment Variables button.
• On either list, look for Path and open it up with Edit...
• You should see one list. Press the New button
• Paste in the path of your build.

Now all you have to do is type noema --model YOUR_CHOSEN_MODEL anywhere in the cmd.

Some useful commands

In the agent chatbot, you can override the persona to whatever you want for the next response. /overridepersona and press Enter.

Also, you can view the previous persona with /persona.

Now with all that out of the way:

How does it work

Noema uses a simple pipeline:

On every prompt:

• A random archetype is selected from a list
• The archetype is turned into behavioural instruction
• That instruction is then used on a simple AI agent.

That's it.

Persona archetypes

Noema starts with a large collection of 1000* persona archetypes

Examples include:

• The Absurdist
• The Romantic
• The Sentimentalist
• The Pattern Poet
• The Material Fetishist
• The Detail Hoarder

(p.s. I generated them all)

These act as seeds from which a more detailed behavioural perspective is generated.

Before Noema starts generating a response, one of these persona archetypes are picked at random. Then behavioural instructions get generated.

Dynamic behavioural instructions

Using the archetype from above, Noema asks the local language model (LLM) to turn it into a behavioural persona.

This persona is then combined with a set of fixed guardrails. (I call it guardrails but it just makes sure the LLM doesn't make unintended results)

These guardrails ensure that the persona remains a perspective rather than becoming a rigid procedure.

The underlying model stays the same

Noema doesn't switch to a different model for each persona. The same model receives the same general task, while the generated persona changes.

Figure A

This makes it possible to investigate whether the persona itself contributes to output diversity.

Why?

LLMs can produce diverse responses, but their outputs often show recurring patterns and stylistic convergence.

i.e. echoes, quantum, pulse, synergy

AI really loves to say words like this. They're a thirsty quantum-chan.

Anyways...

Noema explores whether introducing different behavioural perspectives can push the model into different creative paths. There are promising studies to back up this hypothesis.

Current implementation

Noema is currently a local C# application using .NET 10, Ollama, Spectre.Console and a locally stored list of persona archetypes.

The initial developer, IGE (Isaac Shin) will most likely not contribute to major updates to the software, as he is a game developer who had a cool idea.

Related studies

Why default AI output is narrow

• Kirk et al., "Understanding the Effects of RLHF on LLM Generalisation and Diversity," ICLR 2024 - arxiv.org/abs/2310.06452. RLHF significantly reduced output diversity compared with SFT in their experiments.

• Mohammadi, "Creativity Has Left the Chat: The Price of Debiasing Language Models," 2024 - arxiv.org/abs/2406.05587. In Llama-2 experiments, aligned models showed lower diversity and clustering/attractor-state behaviour.

• Karouzos et al., "Where does output diversity collapse in post-training?," 2026 - arxiv.org/abs/2604.16027. Across several post-training lineages and tasks, output diversity collapsed at different stages depending on training-data composition.

• Doshi & Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content," Science Advances 10(28), 2024 - doi.org/10.1126/sciadv.adn5290. A key empirical motivation for research into AI-assisted creative homogenization.

• Anderson, Shah & Kreminski, "Homogenization Effects of Large Language Models on Human Creative Ideation," Creativity & Cognition 2024 - arxiv.org/abs/2402.01536. Different people brainstorming with ChatGPT converge toward each other.

Prior work on persona-driven generation

• Chan et al., "Scaling Synthetic Data Creation with 1,000,000,000 Personas" (Persona Hub), 2024 - arxiv.org/abs/2406.20094.

• Wan & Kalman, "Diverse AI Personas Can Mitigate the Homogenization Effect in Human-AI Collaborative Ideation," Computers in Human Behavior: Artificial Humans, 2026 - arxiv.org/abs/2504.13868. One of the closest existing studies to Noema's core hypothesis as of 13th August 2026: 10 personas, embedding-confirmed diversity, tested against a human baseline.