New method improves LLM verification with continuous scoring

Original: This new paper from Stanford, Berkeley and NVIDIA feels like a massive piece of the puzzle when it comes to test time compute.

Source: x.com ↗