Event · Release
NVIDIA
– https://arxiv.org/abs/2611.09891
英伟达发布论文《CPU-ARC: Agentic Reasoning Caching for on-CPU Reasoning Reuse》,提出将模型推理过程转换为可在CPU上复用的代码。
1 reports2 sources
Reports 1
Rohan Paul (@rohanpaul_ai)
Machine translation
CPU-ARC study proposes reusing repeated reasoning as CPU code
A paper on CPU-ARC proposes having Claude Opus 4.5 turn a reasoning trace into Python functions with variable inputs. The functions then run on a CPU, with a fine-tuned 1B model used only for small text tasks such as extracting values. The cited summary says the cache matched or beat the source model’s accuracy on seven of eight math and science benchmarks; those results do not establish that it works for every repeated task.
SourceRohan Paul (@rohanpaul_ai) ↗Rohan Paul (@rohanpaul_ai) ↗