An invariant is a statement that is works as limitation and remains true under the condition you apply. Why it matters and how to use it? You can think about invariant as a verification step, you can think about invariant as a weak for of an axiom or even as a weak for of a theorem, it's something that you believe is true and you tested it and it seems convincing that it's true.
For instance: 2+2=4, is an invariant. What property of it is important? Well, you and I know that it's true. we don't need to communicate to give the same answer for the question.
Another example: distance between moon and eatrh is 384 thousands kilometers.
How we can use invariants in our system? We can use them to verify that LLM produces valid result, since we know that the result of an invariant has to be some value, then we can give an LLM a query and expect to have a result that matches expectation. For instance, asking an LLM what 2+2 is equal to, we expect the answer is 4, if LLM gives another response then we can say that the answer if wrong and we can ask LLM correct its answer.
we all need to learn about it (youtube is waiting you)
what we can try in particular and probably it could be cool IDEA to try to apply Zero Knowledge proofs (if possible) as an approach to proof that LLMs' answer is legit. Here you go
Learn about mathematical proofs and invariants.
Invariants and why we need them?
An invariant is a statement that is works as limitation and remains true under the condition you apply. Why it matters and how to use it? You can think about invariant as a verification step, you can think about invariant as a weak for of an axiom or even as a weak for of a theorem, it's something that you believe is true and you tested it and it seems convincing that it's true.
For instance: 2+2=4, is an invariant. What property of it is important? Well, you and I know that it's true. we don't need to communicate to give the same answer for the question. Another example: distance between moon and eatrh is 384 thousands kilometers.
How we can use invariants in our system? We can use them to verify that LLM produces valid result, since we know that the result of an invariant has to be some value, then we can give an LLM a query and expect to have a result that matches expectation. For instance, asking an LLM what 2+2 is equal to, we expect the answer is 4, if LLM gives another response then we can say that the answer if wrong and we can ask LLM correct its answer.
Proofs: