DETECTING WHEN LLMS ARE UNCERTAIN

Summary from the TLDR Web Dev newsletter

Reasoning in LLMs can be improved by utilizing adaptive sampling techniques that take into account the model's uncertainty. A project called Entropix measures uncertainty using entropy and varentropy, which quantify the distribution of the predicted tokens. It then uses different sampling methods based on these metrics, including argmax sampling for high confidence, branching for multiple possibilities, and inserting "thinking" tokens for low confidence.