This is a modest start on an important direction for AI Alignment work; which is, as the authors observe, commonly comprised of tasks which are not readily empirically verifiable and not easily mathematically modeled - so it's hard to get at with normal RL techniques.
I find the ACCoRD benchmark the most interesting, because you could theoretically scale it up from the baseline mode of testing two instances of the same model for their `P(A) ≥ P(A&B)` respectively, you could do `P(A)≥ P(A&B) && P(A) ≥ P(A&C) && P(A&B) ≥ P(A&B&C) && P(A&C) ≥ P(A&B&C) ...` etc
i.e. a swarm of model instances could be collectively measured for consistency for even more confidence, right?
At any rate, even the basic idea of measuring a model for consistency in beliefs improves our ability to bound the amount of trust we can put on it with introspection methods.
Comments
This is a modest start on an important direction for AI Alignment work; which is, as the authors observe, commonly comprised of tasks which are not readily empirically verifiable and not easily mathematically modeled - so it's hard to get at with normal RL techniques.
I find the ACCoRD benchmark the most interesting, because you could theoretically scale it up from the baseline mode of testing two instances of the same model for their `P(A) ≥ P(A&B)` respectively, you could do `P(A)≥ P(A&B) && P(A) ≥ P(A&C) && P(A&B) ≥ P(A&B&C) && P(A&C) ≥ P(A&B&C) ...` etc
i.e. a swarm of model instances could be collectively measured for consistency for even more confidence, right?
At any rate, even the basic idea of measuring a model for consistency in beliefs improves our ability to bound the amount of trust we can put on it with introspection methods.