Tag
A new paper shows that scaling inference compute via methods like self-consistency improves LLM accuracy in math and code but fails to improve truthfulness in domains without external verifiers, as model errors are too correlated.