Probe Classifier
An interpretability technique in which a linear classifier is trained on the intermediate-layer activations of a model to detect whether a specific concept is encoded; high accuracy is often taken as evidence for the presence of that concept.
Timeline (last 90 days)
Model interpretability tools (such as attention visualization, probe classifiers and mechanistic interpretability analysis) can currently provide only partial insight and cannot yet give complete behavioral guarantees.
The high accuracy of a probe classifier can be inversely interpreted as 'there exists a set of coordinates that makes the concept linearly separable', and this interpretation is fully compatible with the new paradigm.
Internal Representation refers to the way large language models encode concepts, relationships, and semantics within neural network layers, and must be studied using interpretability techniques such as probe classifiers and activation analysis.
All Facts (3)
Model interpretability tools (such as attention visualization, probe classifiers and mechanistic interpretability analysis) can currently provide only partial insight and cannot yet give complete behavioral guarantees.
50%UnverifiedThe high accuracy of a probe classifier can be inversely interpreted as 'there exists a set of coordinates that makes the concept linearly separable', and this interpretation is fully compatible with the new paradigm.
50%UnverifiedInternal Representation refers to the way large language models encode concepts, relationships, and semantics within neural network layers, and must be studied using interpretability techniques such as probe classifiers and activation analysis.
50%