KLA is a semiconductor process control company. https://www.kla.com/ Rick Wallace is the CEO.

• 135000 employees
• 8.2B of revenue
• 72-300 tools
• 15% of revenue in R&D

Their main business is in automatically inspecting chips and wafers in time.

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 \(A\).

• hard concrete
• slimming
• zero out
• IG
• activation norm
• random

A KS test is a hypothesis test that measures if two groups of samples are drawn from the same distribution.