This is a very interesting paper. I had some questions and comments about the setup and results:
1.a. In the basic model, the probability of auditing only depends on the prompt, but in the LLM case, the auditor now has access the solver's proposed solution. This seems fundamentally different: now, the solver may exhibit some maladapted behavior to exploit the auditor's sense of when to audit. Do you have any intuition as to how this equilibria might be affected by this?
1.b. In the setting of the near-future where agents are used to make impactful decisions... (read more)
This is a very interesting paper. I had some questions and comments about the setup and results:
1.a. In the basic model, the probability of auditing only depends on the prompt, but in the LLM case, the auditor now has access the solver's proposed solution. This seems fundamentally different: now, the solver may exhibit some maladapted behavior to exploit the auditor's sense of when to audit. Do you have any intuition as to how this equilibria might be affected by this?
1.b. In the setting of the near-future where agents are used to make impactful decisions... (read more)