You Can Make Any Model Jev, But There’s a Catch

I thought the first example would be easy. I asked a local model about releasing research code and data, got an almost-certain answer, and had to revisit the wording of my own question.

I was trying to give the model a smaller job: read a text and choose from defined answers, so ordinary code could decide what happened next.

Jev is built for that. AnyJev promises a similar interface around an open language model without retraining it. And a regular generative model can answer a bounded question too.

AnyJev gave me the kind of answer my code could use. The catch showed up when I asked what would happen if the code trusted those probabilities.

In my earlier Jev article, I worked through issue triage and source checking. Nokia’s AnyJev gives me a way to explore the same idea with open weights. I tried it on scientific claims so we could follow one question through the different approaches, including where the appealing interface stopped helping.

First, let the language model answer

Imagine you have a claim and a paper’s abstract. You want to know whether the abstract supplies evidence for that claim. For a literature tool, the result might decide which papers need another look.