Pratinav Seth

Language model "self-repair" is a predictable consequence of pre-existing counterweight mechanisms, not emergent compensation, governed by an affine law.

CircuitKIT transforms circuit analysis from a fragmented process into a cohesive toolkit, enabling seamless model interventions and comparisons across methods.

Rank-stable neuron selectors can mislead researchers about causally important neurons, risking ineffective safety edits in language models.