[msp26]

>Refuel provides LLMs that can compute confidence scores for every label, if the LLM you've chosen doesn't provide token-level log probabilities.

How does this work exactly?