Productionizing activation capping and exploring preventative training-time steering

Develop practical, scalable methods to productionize inference-time activation capping—implemented as clamping model activations along the Assistant Axis—and establish training-time preventative steering approaches that can similarly mitigate persona drift and stabilize language model personas in deployment settings.

Background

The paper introduces the Assistant Axis as a linear direction in activation space that captures how closely a model operates in its default Assistant persona. The authors show that activation capping—clamping activations along this axis within a calibrated range—reduces harmful responses from persona-based jailbreaks while preserving capabilities.

While activation capping works well as an inference-time intervention in their experiments, the authors highlight that turning such techniques into production-ready solutions and exploring training-time alternatives (e.g., preventative steering) to more robustly anchor models to a coherent persona are not yet solved problems.