PORA: Learning When More Reasoning Pays Off for LLM Agents

Forum topic · 小凯 · 2026-10-22

Summary

PORA (Paired Outcome-guided Reasoning Allocation) is a method for LLM agents that decides, before each action, whether to use short or long reasoning when the total reasoning budget is limited. Longer reasoning can prevent later errors but consumes budget needed for future decisions. PORA learns this trade-off by completing the task twice from the same interaction state during training—once with short reasoning and once with long reasoning—and comparing final task quality and total cost to estimate the downstream value of longer reasoning. At inference time, a controller weighs this estimate against a learned price reflecting remaining budget pressure, selecting short or long reasoning without running both alternatives. Evaluated on application, memory, code, retail, and math tasks, PORA achieves strong results across all settings, outperforming the strongest prior method by 17.4–32.9 percentage points on the BFCL memory task. Within the same agent, it matches or exceeds always-long reasoning on code and math tasks while reducing agent tokens by 50.6% and model requests by 31.2%. Paper: arXiv 2610.27119 by Bo Yan, Qi Zhang, Zhen Tan, Tianlong Chen, and Song Wang.

Paper Overview

Field: Machine Learning Authors: Bo Yan, Qi Zhang, Zhen Tan, Tianlong Chen, Song Wang Published: 2026-10-22 arXiv: 2610.27119

Abstract

LLM agents perform tasks over multiple steps and revise their actions based on feedback. As the total reasoning budget is often limited and must support future steps, the agent must choose between short and long reasoning before each action. Longer reasoning may improve the final task outcome by preventing later errors, but it also uses budget capacity that may be needed for future decisions. This creates two challenges: estimating the downstream value of long reasoning and deciding when that value is worth using the remaining budget.

The authors propose PORA (Paired Outcome-guided Reasoning Allocation), which learns this decision by comparing short and long reasoning from the same interaction state. During training, PORA completes the task twice from each selected state—once with a short reasoning allocation and once with a long one—comparing final task quality and total cost to estimate when longer reasoning pays off. At inference time, a controller weighs this estimate against a learned price reflecting remaining budget pressure and selects short or long reasoning, without running both alternatives simultaneously.

Results

Evaluated on application, memory, code, retail, and math tasks, achieving strong results across all settings.
On the BFCL memory task, PORA outperforms the strongest existing method by 17.4–32.9 percentage points.
Within the same agent, it matches or exceeds always-long reasoning on code and math tasks while reducing agent tokens by 50.6% and model requests by 31.2%.

---

*Auto-collected on 2026-10-22*