Mechanistic
interpretability
We study how generative models represent concepts internally. Our goal is to steer or suppress a specific concept in the model’s output while preserving its other capabilities.
Trace a concept through the model
Find a feature
Identify a pattern associated with a concept.
Intervene
Adjust the feature to test its effect.
Before intervention
After intervention
Compare outputs
Measure the target effect. Check other capabilities.
Illustrative activation view.
Research priorities
01
Identify candidate features
Look for internal features associated with a concept and check whether the association holds across prompts.
02
Test causal interventions
Intervene on a candidate feature and measure the effect on the target behavior.
03
Check for unintended changes
For video, check whether scene content, motion or sharpness changes on prompts with and without the target concept.
