Glossary · Term

concept clamping

← all terms

Definition

Plain language

Forcing a model's internal state to a chosen value along one direction, to test what that direction controls.

As stated in the literature

An interpretability intervention that projects out the existing component along a target direction in activation space and writes in a fixed value, isolating that direction's causal effect more cleanly than additive steering.

Why it matters: It lets researchers test what an internal direction actually controls, turning guesses about a model's insides into checkable cause-and-effect.

For example, researchers can pin the internal 'bridge' direction to a high value and watch whether the model suddenly starts talking about bridges.

Heard on the show

“So they do a second experiment that's stricter, called concept clamping.”
Episode 018 — Language Models Compute the Rational Move, Then Override It

Mentioned in 1 episode

  1. 018
    Language Models Compute the Rational Move, Then Override It

Related terms