Reproduce a Mechanistic Interpretability Result on a Small Transformer
Overview
What this challenge is about.
Pick a published mechanistic-interpretability paper that operates on a small (under 1 billion parameter) open-source transformer (e.g., GPT-2 small, Pythia 70M). Set up the environment, reproduce the headline finding, and run at least 2 follow-up experiments that vary one factor (model size, task, layer). Document everything in a 6-page reproduction report. Include circuit diagrams or attention-pattern visualizations. Be explicit about what you reproduced cleanly and what required interpretation. Add a 1-page reflection on what the finding does and does not tell us about alignment.
The Brief
What you'll do, and what you'll demonstrate.
Reproduce a published mechanistic-interpretability result and extend it with at least 2 follow-up experiments.
Earning criteria — what you'll demonstrate
- Reproduce a published mechanistic-interpretability finding
- Use standard tooling (TransformerLens or equivalent) to probe a small model
- Design follow-up experiments that vary a single factor
- Reason honestly about what an interpretability finding does and does not show
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
ML Researcher
Mechanistic-interpretability reproduction is the leading hiring signal for junior interpretability ML researchers at top safety labs.
This challenge sharpens
- mechanistic-interpretability
- transformer-internals
- research-writing
AI Safety Researcher
Interpretability work sits at the heart of modern AI safety research; this challenge builds the exact skill stack.
This challenge sharpens
- mechanistic-interpretability
- alignment-research
- experiment-design
Research Scientist
Designing follow-up experiments that vary one factor at a time is the research scientist's quality bar.
This challenge sharpens
- experiment-design
- pytorch
- research-writing