controllable
We want \(P(Y|X) = p\), for a specific \(p\) that we specify.
fine-grained control
ideally, instead of optimizing over entire expected values, we want to tune specific utputs
Success in Editing
Say we edited some \(M\), specifying a viper is a vertebrate.
Ideally, this should also edit the other related information:
- \(P\) (paraphrases)j: viper and vertebrates
- \(E\) (logical entailments): a viper has a brain
And we shouldn’t touch:
- \(R\) (other stuff): Chile is a country
- \(LN\) (local neural data): a viper is venomous
Hypernetwork Weight Editing’s Drawbacks
- harder to fix errors than creating them
- harder to retain preformance on local data than random data
- hander to generalize to entailed data than paraphrases
- Updates improves consistency
Information Deletion
- “deleting information” from LLMs is undefined
- RLHF, SFT, etc. HIDES rather than ddeleting
- this can be framed as model editing
High Level Approach
- notice threat information
- attempt to “delete it”
- evaluate the deletion
- try to extract the threat information again
- loop
We formalize this by saying, for some adversarial \(A\) to question \(Q\), we hope that the candidate output set \(C\) of size \(B\) all don’t contain \(A\).
Formal guarantees don’t work very well in LLMWorld.
Ideally, we balance attack success and the damage to other aspects from the model.
Supervision Gap Recovered
Measuring the ratio between the rate of success of “easy” supervision data over “hard” supervisiation data.