Meta 论文提出 Meta-Reasoning Agent，让 worker 执行任务、独立 controller 决定下一步，在最大预算下于全部 12 项对比测试中胜过无 manager 的同款智能体。用 GPT-5.5 跑代码基准时，预算增至 3 倍，得分从 64.1% 升至 71.5%，而对照智能体停滞在 64% 附近。
New Meta paper shows that long-running agents keep getting better with more compute when a separate manager decides how to spend it.
More compute gives an agent more choices: what to try next, what to trust, when to stop. Many agents make those calls on the fly, so extra budget can go to waste.
Their fix hands those calls to a separate manager that thinks them through, while workers do the actual task.
They built a Meta-Reasoning Agent: workers do the task, and a separate controller decides what comes next. It keeps a short progress summary, weighs options against the remaining budget, and picks which past results each worker sees.
At the largest budget, this setup beat an otherwise identical agent without the manager in all 12 head-to-head tests. With GPT-5.5 on a coding benchmark, tripling the budget lifted its score from 64.1% to 71.5%, while the other agent stalled near 64%.
For long-running agents, don't just add compute: spend some on a manager that decides where the rest goes.
What if an agent could decide how to structure its own computation? Agentic Meta-Reasoning: Let the model decide what to explore, what to build on, what to verify, and where to spend its next unit of compute: effectively constructing its own computational graph as it reasons. 🧵 Paras Dahal, @anton_bakhtin , @TacoCohen , Zhengxing Chen, Carole-Jean Wu, Rob Fergus, Scott Yih, @syhw , @rsalakhu , @prfsanjeevarora , @jaseweston在 X 查看被引用的帖子
来源：Rohan Paul · x.com