Skip to content
Tracks @XLab

Spillway design vs. inoculation prompting

Spillway design vs. inoculation prompting

Inoculation prompting

mechanism: prompting

  • requires the AI to attend to the inoculation prompt during training
  • may degrade over the course of training — since the prompt always recommends reward hacking, RL may favor ignoring it
  • some specific circuits that produced reward hacking might not have attended to that part of the prompt

Spillway design

mechanism: pre-RL priors

  • pre-RL priors might shape generalization in a way that survives throughout RL
  • might directly rewrite existing associations with reward hacking (like power-seeking)

Embed this demo

Use the chrome-less embed view inside an iframe:

<iframe src="/demos/spillway-vs-inoculation/embed" width="100%" height="360" style="border:0"></iframe>