Non-linear Interventions on Large Language Models
For researchers and practitioners seeking to understand and control LLM behavior, this provides a more general intervention framework that can handle non-linear features, addressing a key limitation of existing linear methods.
This work extends intervention methods for LLMs beyond linear representations to non-linear features, enabling more precise steering of model behavior. In refusal bypass steering, their method outperforms linear baselines by intervening on a non-linear feature governing refusal.
Intervention is one of the most representative and widely used methods for understanding the internal representations of large language models (LLMs). However, existing intervention methods are confined to linear interventions grounded in the Linear Representation Hypothesis, leaving features encoded along non-linear manifolds beyond their reach. In this work, we introduce a general formulation of intervention that extends naturally to non-linearly represented features, together with a learning procedure that further enables intervention on implicit features lacking a direct output signature. We validate our framework on refusal bypass steering, where it steers the model more precisely than linear baselines by intervening on a non-linear feature governing refusal.