How inference-time scaling, self-consistency, verifiers, and reasoning budgets improve LLM quality, and when not to use them due to high costs.