This reminds me of the Kernel Language Entropy (KLE) paper (arxiv.org/pdf/2405.20003). They use unit trace positive semidefinite kernels to encode the semantic similarities of LLM outputs and quantify uncertainty using the von Neumann entropy.

If you're excited about building the future of automated research, you'll love the #AI4Science community on alphaXiv. Lots of interesting discussions and talks that you don't wanna miss! Link to join: alphaxiv.org/channels/towar…