There has been work that has pushed performance too, like tiny recursive models. Applying LLMs in recursive loops also improves output, so efficiency improvements make this viable, which can count as improvements in performance.