Tag archive
New work targets the compute burned when a model produces confident, plausible reasoning on problems it cannot solve, with one method training models to refuse instead of reasoning and another cutting total token use by up to 68.5 percent by realloca
A controlled experiment counting every token spent found that seven popular self-improvement methods - self-refinement, reflection, debate, best-of-N selection - never reliably beat the simplest baseline of generating several answers and taking the m
A new analysis shows that sampling many answers from an AI and picking the most common one hits a hard ceiling because the samples are correlated, not independent, so thousands of extra tries can be worth only a couple of genuinely new ones.
Controlled experiments on constraint satisfaction problems. MCTS beats best-of-N only when blind sampling hits a ceiling and the verifier provides a gradient.