LLM agents solve multi‑step tasks by revising actions based on feedback, yet the total reasoning budget is limited. Before each action the agent must decide between short and long reasoning. Longer reasoning can prevent later mistakes and improve the final outcome, but it consumes budget that may be needed for future decisions, creating two challenges: estimating the downstream value of long reasoning and deciding when that value justifies the remaining budget.
We introduce Paired Outcome‑guided Reasoning Allocation (PORA), which learns this decision by pairing short and long reasoning runs from the same interaction state. During training, \method{} completes the task twice from each selected state—once with a short allowance and once with a long allowance—and compares their final quality and total processing cost to estimate when longer reasoning is beneficial. At inference time, a controller weighs this estimate against learned prices for remaining‑budget pressure and selects either short or long reasoning without executing both alternatives.
Across application, memory, code, retail, and mathematics benchmarks, PORA delivers strong results in every setting. On the BFCL memory task it outperforms the strongest prior method by 17.4%–32.9% absolute. Within the same agent, it matches or exceeds always‑long reasoning on code and math while using 50.6% fewer actor tokens and 31.2% fewer model requests, respectively.
Review