Evaluation
Probing (interpretability)
2016ActiveUpdated: 22 September 2026Published
Key
innovation
Reading encoded information from activations via a simple, trained classifier.
Category
Evaluation
Abstraction level
Pattern
Operation level
InferenceEvaluation (runtime)
Use cases
LLM representation analysisKnowledge/bias detectionInterpretability research
How it works
Freeze the model, collect activations for inputs, and train a simple classifier predicting the target feature from those activations.
Problem solved
We need to determine what information is encoded in a model's hidden representations.