Oliver Eberle Max Planck School of Cognition Center for L Learning

Sparse interventions can activate complex task behaviors in language models by targeting just 0.01% of neurons, revealing hidden nonlinearities in model dynamics.