Researchers led by Qi Zhang released PORA, a technique for LLM agents that learns when longer reasoning improves results. During training the system runs the same step twice—once with a short reasoning limit and once with a longer one—to see which yields a better final outcome. At run time a controller uses that learned estimate, plus the remaining computation budget, to choose the cheaper option without running both. Tests on memory, code, retail and math tasks showed up to 32.9% higher scores on a memory benchmark and up to 31% fewer model calls on math tasks.
Why it matters
It lets developers run smarter AI agents that use less cloud compute while keeping performance, lowering costs for services that rely on multi‑step reasoning.
