Adding Static and Dynamic Personas to Ideation: Bringing Human Research Vibes to AI Scientist v2

[Mirmix]

Hey everyone! 👋

First off, huge thanks for the nice work on AI Scientist v2—it’s such a tempting project! I’ve been playing with it on the side (with limited time and compute, as one does 😅) and wanted to share a simple idea (and hopefully start some discussion) I’ve been prototyping: a Persona System with static and dynamic personas to make the ideation process feel more like how we humans tackle research. Since LLMs are already emulating us, I thought this could be a fun way to add that human spark to the AI Scientist’s brainstorming. It’s early days, but I’d love to kick off a discussion to get your thoughts or see if anyone’s tried something similar!

The Story Behind It

When we humans dive into research (including ideation), we start with big, bold dreams—think “change the world” energy. But as we iterate, talk with mentors, and get feedback, our ideas evolve. They get sharper, more practical, sometimes totally reshaped. It’s this messy, human process that makes research so exciting, and I wanted to bring that vibe to AI Scientist v2.

So, I’ve cooked up a lightweight persona system to mimic that journey. Picture virtual research team members—it starts with an idiot dreaming big, others poking holes—guiding the AI through a feedback-like loop. I kept it simple due to my time and compute budget, but I think it can be a good starting point.

What’s in the Mix?

I’ve added three modes to the ideation process:

• Baseline Mode: This is the current AI Scientist v2 setup. No personas, just the original system generating and refining ideas without role-specific perspectives. It’s the foundation we’re building on.
• Static Mode: This introduces a fixed set of personas that stay consistent across workshops. They’re like a trusty research crew:
  • Visionary Pioneer: Pushes for bold, paradigm-shifting ideas. (aka Round 1)
  • Creative Strategist: Finds clever, elegant solutions. (aka Round 2)
  • Rigorous Reviewer: Checks for technical soundness and feasibility.
  • Skeptical Critic: Questions assumptions and hunts for flaws.
  • Synthesis Expert: Ties everything into a polished proposal.
• Dynamic Mode: This gets more flexible! Personas adapt to the workshop’s context—its research domain (like healthcare or ML) and themes (like ethics or robotics). For example, a healthcare workshop might get a “Medical Innovator,” while an ML conference might have a “Theoretical ML Guru.” It tailors the brainstorming to the problem. At least this is a final goal, not exactly the current implementation.

Why This Feels Right

Since LLMs are mimicking how we think (learned on human-generated data), adding personas feels like a natural way to make ideation more human-like. Static personas provide a reliable structure, like a familiar team, while dynamic personas adapt to the context, like researchers pivoting for a new project. Having variance in reflection personas could help the AI Scientist generate ideas that start bold and get refined, just like we do than doing everything all alone and many many times without outside feedback.

One idea I haven’t fully explored (due to compute limits) is generating dynamic personas with an LLM for richer, character-like roles. This could make them even more vibrant but would bump up token usage, may slightly increase the cost per paper (already ~$15–$20 per run, as noted in the paper).

How I’ve Been Testing It

I’ve built a simple prototype in my fork: Mirmix/AI-Scientist-v2-personas. The key bits are:

• reflection_personas.py: Handles persona logic, including theme extraction and domain detection for dynamic personas.
• perform_ideation_temp_free.py: Updated to support baseline, static, and dynamic modes, with options to save personas for analysis.

I ran some small experiments (2 idea generations, 6 reflection iterations—tight budget!) at ai_scientist/ideas. So far, both static and dynamic personas seem to produce more evolved experimental suggestions than the baseline. But again, more experiments need to be carried out to get better insights.

Let’s Brainstorm Together!

I know this is a baby idea—very simplistic due to my side-project constraints—but I’m excited to get your thoughts! Could static and dynamic personas make the AI Scientist’s ideas more creative or robust? Here are some questions to spark discussion:

• Do you think this human-inspired approach could lead to better research proposals?
• Has anyone tried something similar with AI Scientist? What worked or didn’t?
• How could we balance richer personas (e.g., LLM-generated) with keeping costs reasonable?

You can check README_PERSONA.md for setup details. I’d love to hear your feedback, wild ideas, or if you’ve tested anything like this before. Let’s make the AI Scientist’s brainstorming as vibrant as a human research lab! 😄 And sorry for being cheesy.

Cheers,
A fan