LLM-as-a-Verifier: an open engine that scores answers, not trust-me-bro outputs
A general-purpose, open-source framework that stops treating an LLM's word as final: it runs fine-grained verification over candidates, scoring each answer on criteria rather than accepting the model's assertion. The trick is averaging over the full logprob distribution of score tokens plus scaled repeated evaluation. Selection beats Pass@1 even when the verifier judges its own model's rollouts.