CPU-ARC Paper: Reusing Reasoning Caches on CPU at 16-291x Lower Cost

rohanpaul_ai · x · 2026-11-22

The arXiv paper "CPU-ARC: Agentic Reasoning Caching for on-CPU Reasoning Reuse" (Peter Belcak, Greg Heinrich, Saurav Muralidharan, Shizhe Diao, Pavlo Molchanov) argues LLMs conflate task-solving with solution execution, forcing them to re-solve every problem from scratch.

It introduces agentic reasoning caching, which builds robust CPU-executable agentic reasoning caches (ARCs) from LLM reasoning traces. ARCs perform on par with frontier LLMs across difficult math and science tasks, are fully explainable, continually trainable, robust to input perturbations, and 16-291x cheaper to invoke than the source model. The tech report runs 28 pages with 9 figures and 11 tables.

Related event: NVIDIA's CPU-ARC Caches LLM Reasoning as Code, Cutting Costs Up to 291x(2 posts)→

Original post →

More from Infra

Infra channel →