Definition

A linear probe is a linear classifier trained on the activations of layer to detect a hidden property of the input: truthfulness vs. lying, factual correctness vs. hallucination, sentiment, topic, bias.

Probes work because the residual stream is additive, so representations are approximately linear. Example: probes on Chess-GPT, trained only on move text, recover the board state and the player’s Elo.

Appears in