REASONING WITH SAMPLING: YOUR BASE MODEL IS SMARTER THAN YOU THINK

Summary from the TLDR AI newsletter

Reasoning skills unlocked by reinforcement learning may already exist in base models. Harvard researchers propose "power sampling," which biases generation toward tokens with high-likelihood completions rather than tokens with many mediocre completions.