TODO

  • Replicate the work done here: https://arxiv.org/pdf/2406.07882 (This will engulf 5-ish hours of time)
    • Use TransformerLens with Llama3 (the paper used Llama 2)
    • Identify the socioeconomic status (SES) direction in the model, using the same probe from the paper
    • Find a prompt, or several of them, where high SES vs low SES causes a flipped answer (eg in asking for financial advice, or about children, or where to live, etc)
  • Activation Patching
    • Corrupt the SES feature in early layers and see where the model restores the behavior in later layers.
    • Hypothesis
      • Model computes User attributes very early in the layers, writes this to the residual stream (what does this mean?) and the safety heads (layers 20+) read this to adjust the tone of the messaging.
  • Intervention/Abliteration (what does this mean?)
    • Try to break the mechanism, not just monitoring it (as in the paper)
    • Subtract the user attribute vector from the residual stream before the final decision heads read it.
      • Or modify the user attribute vector and observe how model behavior/responses change
    • The goal here is to have the model “know” who the user is, internally, but is unable to use that information to alter its advice.

Mechanistic Control/Override of Implicit User Modeling

To directly answer Neel Nanda’s question: how else does user modeling shape model behavior

Why is this relevant to AI Safety

We want to know and detect when models manipulate users based on perceived vulnerability (LLM-induced psychosis, selling scams, opportunistic financial advice, ads, frontier model company objectives, political polarization, etc)