Linear probing is one of the simplest and most powerful ways to test what information a neural representation makes linearly readable. Freeze the base model, collect hidden activations, train a linear classifier or regressor on a target variable, and measure whether that variable can be recovered on held-out examples.
In representation learning, mechanistic interpretability and neuroscience, linear probes and diagnostic classifiers test syntax, semantics, concepts, task state and latent variables without retraining the underlying network. Their attraction is their restriction: a linear readout asks whether the target is accessible through a simple boundary or projection.
This Cognitive Art owner explains linear probing from first principles: frozen representations, classifiers and regressors, train/test splits, linear separability, probe capacity, selectivity controls, random-label baselines, layerwise probing, token selection, causal limits and intervention follow-up.
1. The canonical question
Given representation h, can a simple linear map recover target y on examples the probe did not train on? This tests accessibility under a declared readout class.
2. Freeze the base model
If probe training changes the representation, the experiment stops diagnosing what the original network already encoded.
3. Decodability is not use
A variable can be linearly decodable even when downstream computation ignores it. Correlated or redundant information can make a probe successful without establishing mechanism.
4. Linear separability is geometric
A successful classifier means the sampled labelled examples are separable by a hyperplane to the degree measured on held-out data. It does not imply one neuron equals one concept.
5. Layerwise probing
Using the same controlled probe across depth can show when a variable becomes more linearly accessible, while changes can also reflect rotation, compression or nonlinear recoding.
6. Probe capacity
A linear probe supports a simplicity claim. A high-capacity nonlinear classifier can recover information in forms that ordinary downstream components may not cheaply access.
7. Controls
Random labels, matched control tasks, balanced metrics and strict train-test separation are needed to distinguish representation structure from probe memorization or leakage.
8. Token choice
Final token, subject token and pooled sequence states answer different questions. Token selection belongs in the hypothesis.
9. Probe weights are not automatically native features
The classifier normal is optimized for discrimination. Steering along it is an experiment, not proof that the model naturally uses that exact direction.
10. Causal follow-up
Use Activation Patching and Causal Tracing when the claim changes from readability to functional use.
11. RSA boundary
The RSA owner compares relational geometry without fitting a target readout. Linear probing asks whether one declared variable is accessible through a simple learned readout.
12. Failure signatures
Leakage, class imbalance, layer cherry-picking, probe overcapacity, causal overclaim and treating probe weights as ground-truth features are the main failure modes.
13. Evaluation blueprint
Predeclare target and split; freeze the model; normalize consistently; regularize the probe; tune on validation data; report held-out metrics and controls; repeat across seeds; test causal use separately.
World Return
A rigorous probe result says exactly what is readable, from which activation, by what class of readout, under which held-out distribution. It does not silently upgrade readability into mechanism.