AI Safety Evals Automated Judges with Personas

🤖AI Transparency Statement: This was fully written by me, and has not gone through any review with AI.

Introduction

This project builds off my early assessment of Automated Auditing Tool’s Place in Alignment Evaluations, specifically what I saw as a risk in the variability that can come from LLM responses, and how using multiple judges might help improve the trustworthiness of the judging. After writing that, I found that the idea of multiple judges had already been written about, with more research ongoing, often studying options like pairwise (one judge argues for the, one judge argues against, a third judge picks the winner). But, what I haven’t observed as much work on yet is the effect of assigning personas to judges.

This project is an opportunity for me to work with my first AI safety automated evals, and also commonly used tools like Inspect and Petri.

Initially, I was going to pick this up solo, but then connected with Daniel Alexander at EAGxBerkeley who had already created some evals research projects and was looking for someone to collaborate with.

Project Stage

Last Updated: September 11th, 2026

Ideation: Daniel has already independently run the first rounds of Haiku Judge Evals where he first observed self-preference, then changed the judging setup which effectively eliminated it. I brought the idea of applying personas back up as another way to see how judging might change. We’ve chatted to flesh out the idea, and have now exchanged an initial proposal with personas, how they might be assessed, and what our next step might be.

I’ll work to keep this page up to date as the project develops, but wanted to create it as a stub to show some of the projects I’m thinking about for Fall 2026.

Overall Project Idea and Hypothesis

We’re combining Daniel’s work on judging preferences on Haiku’s with my base hypothesis that assigning personas and potentially later having multiple judges with different personas work together as a counsel may generate more trustworthy automated evals judging results.

The idea follows my main criticism of single judges and how variable any LLM response can be. My first hypothesis is that at least understanding and logging the persona an evals judge has chosen is important, if not explicitly assigning it one. This comes from my work in industry where we would often have 3 skillsets analyzing the same problem and potential solutions. As product I would advocate for the user and org growth goals, engineering would advocate for simpler and robust technical solutions, and data science would advocate for statistically sound methods. We reached the best solutions by evaluating each problem on our own strengths (rather than trying to assess as all personas), then working together to find the best solution. When judging evals, a persona could involve things like professional background, area of study, years of experience, and even things like locale or culture.

Staging will work roughly like this, with more details to come and potentially more sub-stages:

1. If simply requesting the judge to state a chosen persona changes its response.
2. The impact of assigning different personas.
3. Having one response evaluated by multiple assigned personas, then having them align an answer as a counsel.

Some Challenges We’ve Already Flagged

• Picking which values matter as a persona.
• Testing combinations of personas can quickly become a factorial sized problem, exploding the number of samples and the cost of compute.

Project Plan

Not ready to share yet, will likely link out to a markdown once it’s settled.

My Contribution and What I Want to Learn

This is my first true exposure to AI safety evals and the automated evals tooling.

I’m bringing some ideas and perspective from my work from experience in industry, while Daniel is going to anchor the statistical and technical sides of things, and overall lead the project. I’ll help review ideas, phases, and results and offer new ideas where they seem valuable.

One of my AI Safety technical paths may land me within an org performing or supporting evaluations. Getting to work on one directly will provide a lot of good insight on how they are structured and built. It will better inform me if my instincts on LLM judges were right, or if I need to adjust them. It will give me my first hands on exposure to Inspect and Petri, which I expect to seed more ideas to explore if I do pursue building a standardized automated evals platform.

Our Expected Output

Daniel has been sharing his outputs to his Github along with written insights on LinkedIn. Currently we don’t expect the output to rise to the level of warranting a formal paper. We or I will likely publish a less formal summary of the work once completed, and I’ll likely also write a separate reflection piece on what I learned, how my views have changed, etc.