EmergingInterpretability and model internals

Interpretability Evaluation Scientist

Also called Interpretability Evals Researcher

An Interpretability Evaluation Scientist tests whether explanations of model behaviour and discovered internal features are actually faithful. That protects a company from building safety and product decisions on interpretability claims that look convincing but fail under test.

What they do

They take an explanation of a model, such as a circuit, a feature or an attribution, and try to break it. They design interventions on activations and weights, then check whether the explanation predicts what the model does. They build benchmarks with known ground truth, often using models where the mechanism has been planted. They instrument models, run ablations and patching experiments, and compare competing methods on the same footing. They write up where an explanation holds, where it fails and why. They work closely with researchers who produce the explanations.

What a strong hire delivers

In the first months, a strong hire gives the team a clear, shared definition of what faithful means for its own methods. They deliver a working explanation benchmark and an intervention-based evaluation that can be rerun on new models. They show which current findings hold up and which do not. They leave researchers with sharper experimental habits and fewer unchecked claims.

How you can tell they are good

Ask them to take an interpretability result they have seen and explain how they would try to falsify it. Strong people name specific interventions, controls and failure modes without prompting. Ask for a case where their test overturned a popular explanation. Be wary of anyone who only reports methods that worked, or who treats a plausible visualisation as evidence.

Why it is hard to fill

The title barely exists. This work usually sits inside interpretability, safety or evaluation roles, so nobody searches for it by name. Few people combine mechanistic analysis with real experimental rigour. Researchers who build methods often resist attacking them, and evaluation specialists often lack access to model internals. Mixed backgrounds are normal, so judging the work matters more than reading the CV.

Where to find them

Look in interpretability teams at frontier labs and safety-focused organisations, and in academic groups working on mechanistic analysis or causal methods. Many hold titles such as research scientist, research engineer or evaluations researcher. Others come from deep learning research, statistics or experimental sciences where falsification is routine. Open-source interpretability contributors are also worth a close look.

Indicative pay

Not enough published evidence yet to give a fair range for this role, which is common for work this new. We benchmark it for you as part of a search.

Interview questions for an Interpretability Evaluation Scientist

  1. Describe a time you tested whether an explanation of a model was faithful. What did you do and what did you find?

    Look for a concrete experiment, a clear intervention and an honest outcome, ideally including a negative result. Strong answers explain the control and what would have counted as failure. Worry about vague claims or reliance on visual plausibility.

  2. How would you decide whether a discovered feature is genuinely used by the model rather than merely correlated with its behaviour?

    A good answer covers causal interventions such as ablation, patching or steering, plus controls and checks for off-distribution effects. They should discuss what the intervention itself might disturb. Worry if they rely on correlation or activation strength alone.

4 more questions, with what a strong answer covers, are in the Role Kit.

Job description you can post

We are hiring an Interpretability Evaluation Scientist to test whether our explanations of model behaviour and our discovered features are faithful. You will design the experiments and benchmarks that tell us which interpretability findings we can trust and build on.

What you will do

  • Design intervention-based tests for explanations and discovered features
  • Build and maintain an explanation benchmark with known ground truth where possible
  • Run activation analysis, ablation and patching experiments on our models
  • Compare interpretability methods under consistent conditions
  • Try to falsify internal findings and document where they fail
  • Work with researchers to improve methods using evaluation results
  • Report results clearly to technical and non-technical colleagues

What you will bring

  • Demonstrated experimental work testing claims about model internals
  • Strong deep learning fundamentals and hands-on model instrumentation
  • Clear experimental design, including controls and failure criteria
  • Experience with mechanistic interpretability or closely related analysis
  • Evidence of reporting negative results honestly
  • Ability to write clean, reproducible research code

Nice to have

  • Experience building benchmarks or evaluation suites
  • Background in causal inference or statistics
  • Public write-ups or open-source contributions on interpretability

Free to use and adapt. If it saves you time, a link back to this page is appreciated.

Skills that matter

  • Mechanistic interpretabilityAdvancedInvestigate internal model mechanisms
  • Deep learningAdvancedBuild and diagnose trainable neural models
  • Experimental designWorkingChoose comparisons that identify credible effects
  • Model evaluationWorkingMeasure capability and failure modes credibly

Other roles in interpretability and model internals

Hiring nowMechanistic Interpretability ResearcherA mechanistic interpretability researcher works out how a model computes its outputs, using interventions and controlled tests. It matters because a…Hiring nowInterpretability Research EngineerAn Interpretability Research Engineer builds the tools and experiments that let researchers look inside models. They make internals work reliable…EmergingModel Representation AnalystA Model Representation Analyst studies how a model's internal features change with inputs and training, and tests whether those findings hold up. It…

Related roles in other families

Hiring nowMachine Learning Research ScientistLearning theory and statistical machine learningA Machine Learning Research Scientist develops and tests learning methods against explicit hypotheses. They give the company evidence about what…Hiring nowStatistical Learning ScientistLearning theory and statistical machine learningA Statistical Learning Scientist builds prediction methods that show their assumptions and their uncertainty. For a company shipping models into…Hiring nowLearning Theory ResearcherLearning theory and statistical machine learningA Learning Theory Researcher works out what learning algorithms can and cannot do, and proves it. For an AI company, that means firmer ground for…Hiring nowAutomated Machine Learning ResearcherLearning theory and statistical machine learningAn Automated Machine Learning Researcher builds search procedures that choose and tune models within a fixed compute budget. It matters because…

How we wrote this

Role definitions are our own editorial work, checked against occupational classifications, standards, research and employer pages. Titles vary by company, and this is a description of the work, not a job advert.

← Interpretability and model internalsAll new job families

Start a search

Hiring an Interpretability Evaluation Scientist?

Tell us the work. We find the people already doing it, whatever their job title says, and you meet them already interviewed for your role.