The weak-to-strong assumption is that it is easier to eval the result of a task than to generate it. If it is wrong, human can not make a stronger intelligence than us.

Not true. You can easily train a BERT single class classification model without having to train an LLM.

You can distill evaluation ability

not true at all, you could have a efficient cheap model which is generally terrible at most things but has a savant like capacity for categorizing tasks by requirement and difficulty. Even easier when you dont need to support multiple languages and a truly staggering breadth of domains, like a conventional llm does. You could train a really small model to reject out of domain requests and partition the rest, running at a fraction of the cost of a more capable model.

In this paper, they tried a couple different methods for determining how similar the incoming request is to requests that they have scored in their dataset. Actually, one of the best methods they used does not involve using a model at all to evaluate the incoming query (similarity-weighted ranking) although it _does_ use pre-trained embeddings.

Using this, they were able to produce quite good results applying this similarity measurement to unseen queries using a standard benchmark. The leap of faith here is assuming that the same query similarity method will continue to bear fruit when extended to queries that aren't benchmarkable.