Very cool I'm going to try and play with this later. It looks like llm-consortium [0] but with some nice new features, like confidence gating and pluggable verifiers. So, if a response confidence is below a threshold it is eliminated entirely? Is that the gating?
A little more... context: https://jiaweizzhao.github.io/deepconf/
well, that helped understand better. It was a bit dry with just the ref paper. I like the visuals btw.
Nice, I'd love to see this added to the llm-reasoners project [0]. They've got a nice set of reasoning techniques implemented from papers.